cuda: remove linking to cublasLt (#14790)

Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>

.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/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";
   };

.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

.github/workflows/update-ops-docs.yml

@@ -0,0 +1,40 @@
+name: Update Operations Documentation
+
+on:
+    push:
+        paths:
+            - 'docs/ops/**'
+            - 'scripts/create_ops_docs.py'
+    pull_request:
+        paths:
+            - 'docs/ops/**'
+            - 'scripts/create_ops_docs.py'
+
+jobs:
+    update-ops-docs:
+        runs-on: ubuntu-latest
+
+        steps:
+        - name: Checkout repository
+          uses: actions/checkout@v4
+
+        - name: Set up Python
+          uses: actions/setup-python@v5
+          with:
+              python-version: '3.x'
+
+        - name: Generate operations documentation to temporary file
+          run: |
+              mkdir -p /tmp/ops_check
+              ./scripts/create_ops_docs.py /tmp/ops_check/ops.md
+
+        - name: Check if docs/ops.md matches generated version
+          run: |
+              if ! diff -q docs/ops.md /tmp/ops_check/ops.md; then
+                  echo "Operations documentation (docs/ops.md) is not up to date with the backend CSV files."
+                  echo "To fix: run ./scripts/create_ops_docs.py and commit the updated docs/ops.md along with your changes"
+                  echo "Differences found:"
+                  diff docs/ops.md /tmp/ops_check/ops.md || true
+                  exit 1
+              fi
+              echo "Operations documentation is up to date."

CMakePresets.json

@@ -55,6 +55,17 @@
             "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
         }
     },
+    {
+        "name": "x64-linux-gcc", "hidden": true,
+        "cacheVariables": {
+            "CMAKE_C_COMPILER": "gcc",
+            "CMAKE_CXX_COMPILER": "g++"
+        }
+    },
+    { "name": "x64-linux-gcc-debug", "inherits": [ "base", "x64-linux-gcc", "debug" ] },
+    { "name": "x64-linux-gcc-release", "inherits": [ "base", "x64-linux-gcc", "release" ] },
+    { "name": "x64-linux-gcc-reldbg", "inherits": [ "base", "x64-linux-gcc", "reldbg" ] },
+    { "name": "x64-linux-gcc+static-release", "inherits": [ "base", "x64-linux-gcc", "release", "static" ] },
 
     { "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
     { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },

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

@@ -6,9 +6,9 @@
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md)
 
-Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
+LLM inference in C/C++
 
 ## Recent API changes
 
@@ -17,10 +17,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
+- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
+- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
-- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
@@ -134,6 +133,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
+- [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 
 #### Multimodal
 
@@ -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/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/CMakeLists.txt

@@ -86,8 +86,7 @@ if (LLAMA_CURL)
     endif()
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
     include_directories(${CURL_INCLUDE_DIRS})
-    find_library(CURL_LIBRARY curl REQUIRED)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
 endif ()
 
 if (LLAMA_LLGUIDANCE)
@@ -112,13 +111,13 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.20 (+ fix to build on GCC 15):
-        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
+        # v1.0.1:
+        GIT_TAG d795912fedc7d393de740177ea9ea761e7905774
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE
         CONFIGURE_COMMAND ""
-        BUILD_COMMAND cargo build --release
+        BUILD_COMMAND cargo build --release --package llguidance
         INSTALL_COMMAND ""
         BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h
         UPDATE_COMMAND ""

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"),
@@ -3423,5 +3431,34 @@ 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/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
@@ -522,6 +534,7 @@ 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

@@ -300,6 +300,7 @@ class ModelBase:
                             gguf.MODEL_TENSOR.POS_EMBD,
                             gguf.MODEL_TENSOR.TOKEN_TYPES,
                             gguf.MODEL_TENSOR.SSM_CONV1D,
+                            gguf.MODEL_TENSOR.SHORTCONV_CONV,
                             gguf.MODEL_TENSOR.TIME_MIX_FIRST,
                             gguf.MODEL_TENSOR.TIME_MIX_W1,
                             gguf.MODEL_TENSOR.TIME_MIX_W2,
@@ -668,6 +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"
@@ -803,33 +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 == "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 == "b0a6b1c0bd5998ebd9df08611efde34a4ff03faed45ae09c43e6b31ebd4b94cf":
+            # ref: https://huggingface.co/skt/A.X-4.0
+            res = "a.x-4.0"
+        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 == "2085e1638f6c377a0aa4ead21b27bb4cb941bf800df86ed391011769c1758dfb":
+            # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B
+            res = "exaone4"
 
         if res is None:
             logger.warning("\n")
@@ -1072,7 +1088,14 @@ class TextModel(ModelBase):
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
-        special_vocab.chat_template = "rwkv-world"
+        if special_vocab.chat_template is None:
+            template_path = Path(__file__).parent / "models" / "templates" / "llama-cpp-rwkv-world.jinja"
+            if template_path.is_file():
+                with open(template_path, "r", encoding="utf-8") as f:
+                    template = f.read()
+            else:
+                template = "rwkv-world"
+            special_vocab.chat_template = template
         # hack: Add '\n\n' as the EOT token to make it chat normally
         special_vocab._set_special_token("eot", 261)
         # hack: Override these as they have already been set (incorrectly)
@@ -2758,6 +2781,76 @@ 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
@@ -2771,7 +2864,8 @@ 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.")
@@ -2804,6 +2898,93 @@ 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",
@@ -3491,6 +3672,175 @@ class PlamoModel(TextModel):
         return [(new_name, data_torch)]
 
 
+@ModelBase.register("Plamo2ForCausalLM", "PLaMo2ForCausalLM")
+class Plamo2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.PLAMO2
+
+    def set_vocab(self):
+        # PLaMo 2 uses a custom tokenizer with a .jsonl file
+        # We need to handle this specially
+        tokenizer_jsonl_path = self.dir_model / "tokenizer.jsonl"
+        tokenizer_config_path = self.dir_model / "tokenizer_config.json"
+
+        if not tokenizer_jsonl_path.is_file():
+            raise FileNotFoundError(f"PLaMo 2 tokenizer file not found: {tokenizer_jsonl_path}")
+
+        # Load tokenizer config
+        with open(tokenizer_config_path, 'r', encoding='utf-8') as f:
+            tokenizer_config = json.load(f)
+
+        # Load tokens from JSONL file (actually a list format)
+        tokens = []
+        scores = []
+        toktypes = []
+
+        with open(tokenizer_jsonl_path, 'r', encoding='utf-8') as f:
+            for line_num, line in enumerate(f):
+                if line.strip():
+                    token_data = json.loads(line)
+                    # Format: [token, score, type, ?, ?, ?, ?]
+                    token = token_data[0].encode("utf-8")
+                    score = float(token_data[1])
+                    token_type_str = token_data[2] if len(token_data) > 2 else "NORMAL"
+
+                    tokens.append(token)
+                    scores.append(score)
+
+                    # Map token type strings to GGUF token types
+                    if token_type_str == "UNKNOWN":
+                        toktypes.append(gguf.TokenType.UNKNOWN)
+                    elif token_type_str == "CONTROL":
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    elif token_type_str == "BYTE":
+                        toktypes.append(gguf.TokenType.BYTE)
+                    else:
+                        # Check for PLaMo-2 special tokens
+                        token_str = token_data[0]
+                        if token_str.startswith("<|plamo:") and token_str.endswith("|>"):
+                            toktypes.append(gguf.TokenType.CONTROL)
+                        else:
+                            toktypes.append(gguf.TokenType.NORMAL)
+
+        vocab_size = self.hparams["vocab_size"]
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(gguf.TokenType.UNUSED)
+
+        # Use "plamo2" tokenizer type for PLaMo-2's custom Aho-Corasick tokenizer
+        self.gguf_writer.add_tokenizer_model("plamo2")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # Add special tokens from config
+        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] is not None:
+            token_id = tokens.index(tokenizer_config["bos_token"].encode("utf-8"))
+            self.gguf_writer.add_bos_token_id(token_id)
+        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] is not None:
+            token_id = tokens.index(tokenizer_config["eos_token"].encode("utf-8"))
+            self.gguf_writer.add_eos_token_id(token_id)
+        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] is not None:
+            token_id = tokens.index(tokenizer_config["pad_token"].encode("utf-8"))
+            self.gguf_writer.add_pad_token_id(token_id)
+        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] is not None:
+            token_id = tokens.index(tokenizer_config["sep_token"].encode("utf-8"))
+            self.gguf_writer.add_sep_token_id(token_id)
+        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] is not None:
+            token_id = tokens.index(tokenizer_config["unk_token"].encode("utf-8"))
+            self.gguf_writer.add_unk_token_id(token_id)
+
+        # Add <|plamo:op|> as EOT to ensure appropriate end of generation
+        self.gguf_writer.add_eot_token_id(4)
+
+        self.gguf_writer.add_add_space_prefix(False)
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+
+        # Which layers are Mamba layers
+        # PLaMo 2 uses mamba_step to indicate the pattern (e.g., 2 means every other layer)
+        # This logic matches modeling_plamo.py's is_mamba function
+        mamba_step = hparams.get("mamba_step", 2)
+        mamba_enabled = hparams.get("mamba_enabled", True)
+        mamba_layers = []
+
+        if mamba_enabled:
+            for i in range(block_count):
+                if block_count <= (mamba_step // 2):
+                    # use attention in last layer
+                    is_mamba = (i != block_count - 1)
+                else:
+                    is_mamba = (i % mamba_step) != (mamba_step // 2)
+                if is_mamba:
+                    mamba_layers.append(0)
+                else:
+                    mamba_layers.append(hparams.get("num_key_value_heads", 4))
+
+        if mamba_layers:
+            self.gguf_writer.add_head_count_kv(mamba_layers)
+
+        self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 2048))
+        self.gguf_writer.add_embedding_length(hparams.get("hidden_size", 4096))
+        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))
+
+        # Mamba parameters
+        self.gguf_writer.add_ssm_state_size(hparams.get("mamba_d_state", 64))
+        self.gguf_writer.add_ssm_conv_kernel(hparams.get("mamba_d_conv", 4))
+        self.gguf_writer.add_ssm_time_step_rank(hparams.get("mamba_num_heads", 64))
+        intermediate_size = hparams.get("mamba_num_heads", 64) * hparams.get("hidden_size_per_head", 128)
+        self.gguf_writer.add_ssm_inner_size(intermediate_size)
+        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_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if name.endswith(".A_log"):
+            data_torch = -torch.exp(data_torch)
+        elif name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+        elif name.endswith(".dt_norm_weight"):
+            name = name.rpartition(".dt_norm_weight")[0] + ".dt_norm.weight"
+        elif name.endswith(".B_norm_weight"):
+            name = name.rpartition(".B_norm_weight")[0] + ".B_norm.weight"
+        elif name.endswith(".C_norm_weight"):
+            name = name.rpartition(".C_norm_weight")[0] + ".C_norm.weight"
+        elif name.endswith(".k_weight"):
+            name = name.rpartition(".k_weight")[0] + ".k.weight"
+        elif name.endswith(".q_weight"):
+            name = name.rpartition(".q_weight")[0] + ".q.weight"
+        elif name.endswith(".conv1d.weight"):
+            data_torch = torch.squeeze(data_torch)  # remove (, 1, )
+            assert data_torch.ndim == 2
+        elif name.endswith(".pre_mixer_norm.weight"):
+            data_torch += 1.0
+        elif name.endswith(".post_mixer_norm.weight"):
+            data_torch += 1.0 / 5
+        elif name.endswith(".pre_mlp_norm.weight"):
+            data_torch += 1.0
+        elif name.endswith(".post_mlp_norm.weight"):
+            data_torch += 1.0 / (5**1.5)
+        elif name.endswith(".norm.weight"):
+            data_torch += 1.0
+
+        new_name = self.map_tensor_name(name)
+
+        return [(new_name, data_torch)]
+
+
 @ModelBase.register("CodeShellForCausalLM")
 class CodeShellModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CODESHELL
@@ -4887,6 +5237,9 @@ class Mamba2Model(TextModel):
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                 hparams = json.load(f)
         super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+        self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+        self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model
+        self.n_group = self.find_hparam(["n_groups"], optional=True) or 1
 
     def set_vocab(self):
         vocab_size = self.hparams["vocab_size"]
@@ -4909,12 +5262,9 @@ class Mamba2Model(TextModel):
             self._set_vocab_builtin("gpt-neox", vocab_size)
 
     def set_gguf_parameters(self):
-        d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
-        d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model
-        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 128
-        head_dim = self.find_hparam(["mamba_d_head", "head_dim"],    optional=True) or 64
-        n_group = self.find_hparam(["n_groups"],                     optional=True) or 1
+        d_conv  = self.find_hparam(["conv_kernel", "d_conv"],     optional=True) or 4
+        d_state = self.find_hparam(["state_size",  "d_state"],    optional=True) or 128
+        head_dim = self.find_hparam(["mamba_d_head", "head_dim"], optional=True) or 64
 
         rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
 
@@ -4922,19 +5272,19 @@ class Mamba2Model(TextModel):
         # TODO: does this really matter?
         # skip the assertion for FalconH1 Model
         if self.model_arch != gguf.MODEL_ARCH.FALCON_H1:
-            assert d_inner == 2 * d_model
-            assert d_inner % head_dim == 0
+            assert self.d_inner == 2 * self.d_model
+            assert self.d_inner % head_dim == 0
 
         self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default
-        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_embedding_length(self.d_model)
         self.gguf_writer.add_feed_forward_length(0)  # unused, but seemingly required when loading
         self.gguf_writer.add_head_count(0)  # unused, but seemingly required when loading
         self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_ssm_conv_kernel(d_conv)
-        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_inner_size(self.d_inner)
         self.gguf_writer.add_ssm_state_size(d_state)
-        self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
-        self.gguf_writer.add_ssm_group_count(n_group)
+        self.gguf_writer.add_ssm_time_step_rank(self.d_inner // head_dim)
+        self.gguf_writer.add_ssm_group_count(self.n_group)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
         self.gguf_writer.add_file_type(self.ftype)
 
@@ -4959,10 +5309,7 @@ class Mamba2Model(TextModel):
             # (D is also unsqueezed, but for more straightforward broadcast internally)
             data_torch = data_torch.reshape((*data_torch.shape, 1))
         elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
-            d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-            d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * d_model
-            n_group = self.hparams.get("n_groups", 1)
-            data_torch = data_torch.reshape((n_group, d_inner // n_group))
+            data_torch = data_torch.reshape((self.n_group, self.d_inner // self.n_group))
 
         if name.endswith(".A_log"):
             logger.debug("A_log --> A ==> " + new_name)
@@ -4971,6 +5318,123 @@ class Mamba2Model(TextModel):
         yield (new_name, data_torch)
 
 
+@ModelBase.register("JambaForCausalLM")
+class JambaModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.JAMBA
+
+    def get_vocab_base_pre(self, tokenizer) -> str:
+        del tokenizer  # unused
+
+        return "gpt-2"
+
+    def set_vocab(self):
+        if (self.dir_model / "tokenizer.model").is_file():
+            # Using Jamba's tokenizer.json causes errors on model load
+            # (something about "byte not found in vocab"),
+            # but there's a working tokenizer.model
+            self._set_vocab_sentencepiece()
+        else:
+            # Some Jamba models only have a tokenizer.json, which works.
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        d_model = self.find_hparam(["hidden_size", "mamba_d_model"])
+        d_conv  = self.find_hparam(["mamba_d_conv"],  optional=True) or 4
+        d_inner = self.hparams["mamba_expand"] * d_model
+        d_state = self.find_hparam(["mamba_d_state"], optional=True) or 16
+        # ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
+        dt_rank      = self.find_hparam(["mamba_dt_rank"], optional=True) or -(d_model // -16)
+        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-6
+        n_kv_head = self.hparams["num_key_value_heads"]
+        attn_offset = self.hparams["attn_layer_offset"]
+        attn_period = self.hparams["attn_layer_period"]
+        n_kv_vec = [0 for _ in range(attn_offset)] + [
+            n_kv_head if (i - attn_offset) % attn_period == 0 else 0 for i in range(attn_offset, self.block_count)
+        ]
+
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_context_length(self.find_hparam(["max_position_embeddings", "n_ctx"]))
+        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(n_kv_vec)
+        self.gguf_writer.add_ssm_conv_kernel(d_conv)
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(d_state)
+        self.gguf_writer.add_ssm_time_step_rank(dt_rank)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+        self.gguf_writer.add_expert_used_count(self.hparams["num_experts_per_tok"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        # Mini-Jamba
+        name = name.replace(".moe.", ".feed_forward.")
+        if bid is not None:
+            moe_offset = self.hparams["expert_layer_offset"]
+            moe_period = self.hparams["expert_layer_period"]
+
+            if not (bid >= moe_offset and (bid - moe_offset) % moe_period == 0):
+                name = name.replace(".experts.0.", ".")
+
+        # process the experts separately
+        if ".feed_forward.experts." in name:
+            n_experts = self.hparams["num_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+
+                # merge the experts into a single 3d tensor
+                for wid in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.feed_forward.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    # using the same merged name as qwen2moe
+                    merged_name = f"model.layers.{bid}.mlp.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    yield new_name, data_torch
+            return
+
+        new_name = self.map_tensor_name(name)
+
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+
+        if name.endswith(".A_log"):
+            logger.debug("A_log --> A ==> " + new_name)
+            data_torch = -torch.exp(data_torch)
+
+        yield (new_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("CohereForCausalLM")
 class CommandR2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.COMMAND_R
@@ -5439,7 +5903,58 @@ class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
     def set_vocab(self):
-        self._set_vocab_gpt2()
+        try:
+            self._set_vocab_gpt2()
+            return
+        except Exception:
+            pass
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        if tokpre == "kimi-k2":
+            # Build merges list using the approach similar to HunYuanMoE
+            merges = []
+            vocab = {}
+            mergeable_ranks = tokenizer.model._mergeable_ranks
+            for token, rank in mergeable_ranks.items():
+                vocab[QwenModel.token_bytes_to_string(token)] = rank
+                if len(token) == 1:
+                    continue
+                merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+                if len(merged) == 2:
+                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+
+            # Build token list
+            vocab_size = self.hparams["vocab_size"]
+            special_tokens = tokenizer.special_tokens
+            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
+            tokens: list[str] = []
+            toktypes: list[int] = []
+
+            for i in range(vocab_size):
+                if i not in reverse_vocab:
+                    tokens.append(f"[PAD{i}]")
+                    toktypes.append(gguf.TokenType.UNUSED)
+                else:
+                    token = reverse_vocab[i]
+                    tokens.append(token)
+                    if i in special_tokens.values():
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.NORMAL)
+
+            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)
+            self.gguf_writer.add_token_merges(merges)
+
+            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+            special_vocab.add_to_gguf(self.gguf_writer)
+        else:
+            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
 
     def set_gguf_parameters(self):
 
@@ -6268,6 +6783,75 @@ 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"""
@@ -6332,18 +6916,148 @@ class GraniteMoeModel(GraniteModel):
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
             ]
 
+        has_experts = bool(self.hparams.get('num_local_experts'))
+
         if name.endswith("shared_mlp.input_linear.weight"):
             ffn_dim = self.hparams["shared_intermediate_size"]
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
             gate, up = data_torch.split(ffn_dim, dim=-2)
+            if has_experts:
+                return [
+                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
+                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+                ]
             return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), up),
+            ]
+
+        if not has_experts and name.endswith("shared_mlp.output_linear.weight"):
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), data_torch)
             ]
 
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
+class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
+    """GraniteHybrid is a hybrid SSM + Attention model that uses Mamba2 SSM
+    layers and optionally uses MoE w/ a shared expert"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_HYBRID
+    undo_permute = True
+
+    def __init__(self, *args, **kwargs):
+
+        # Hybrid mamba models use a prefix for the mamba-specific params.
+        # TODO: Extend this if the prefix(es) need to be configurable
+        self.hparam_prefixes = ["mamba"]
+
+        super().__init__(*args, **kwargs)
+
+        # Lists of which layers use ssm vs attention
+        self._attn_layers = self.get_attn_layers()
+        self._ssm_layers = [
+            i for i in range(self.block_count)
+            if i not in self._attn_layers
+        ]
+
+        # n_group and d_inner are used during reshape_tensors for mamba2
+        self.d_model = self.find_hparam(["hidden_size", "d_model"])
+        self.n_group = self.find_hparam(["n_groups"])
+        self.d_inner = self.find_hparam(["expand"]) * self.d_model
+
+    def get_attn_layers(self):
+        # Explicit list of layer type names
+        if layer_types := self.hparams.get("layer_types"):
+            return [
+                i for i, typ in enumerate(layer_types)
+                if typ == "attention"
+            ]
+
+        # Layer types indicated by index or period
+        attn_layers = self.hparams.get("attn_layer_indices", [])
+        if not attn_layers:
+            attn_period = self.hparams.get("attn_layer_period")
+            assert attn_period, "Didn't find attn_layer_indices or attn_layer_period"
+            attn_offset = self.hparams.get("attn_layer_offset")
+            assert attn_offset is not None, "No attention layer offset set with attn_layer_period"
+            attn_layers = [
+                i for i in range(self.block_count)
+                if i % attn_period == attn_offset
+            ]
+        return attn_layers
+
+    def find_hparam(self, keys: Iterable[str], *args, **kwargs) -> Any:
+        prefixed = []
+        for pfx in self.hparam_prefixes:
+            prefixed.extend(
+                "_".join([pfx, k])
+                for k in keys
+            )
+        keys = list(keys) + prefixed
+        return Mamba2Model.find_hparam(self, keys, *args, **kwargs)
+
+    def modify_tensors(
+        self, data_torch: Tensor, name: str, bid: int | None
+    ) -> Iterable[tuple[str, Tensor]]:
+        if (
+            name.endswith("block_sparse_moe.input_linear.weight")
+            or "shared_mlp" in name
+        ):
+            return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+
+        # Determine whether this is a mamba layer or an attention layer
+        if bid in self._ssm_layers:
+            return Mamba2Model.modify_tensors(self, data_torch, name, bid)
+        elif bid in self._attn_layers:
+            return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def set_gguf_parameters(self):
+        """This method merges params from both parents and some that are
+        specific to this model. The result is some duplication of how the params
+        get set. The following warnings are expected during conversion:
+
+        WARNING:Duplicated key name 'granitehybrid.attention.head_count_kv'
+        WARNING:Duplicated key name 'granitehybrid.context_length'
+        """
+        GraniteMoeModel.set_gguf_parameters(self)
+
+        ## Mamba mixer params ##
+        self.gguf_writer.add_ssm_conv_kernel(self.find_hparam(["conv_kernel", "d_conv"]))
+        self.gguf_writer.add_ssm_state_size(self.find_hparam(["state_size", "d_state"]))
+        self.gguf_writer.add_ssm_group_count(self.n_group)
+        self.gguf_writer.add_ssm_inner_size(self.d_inner)
+        # NOTE: The mamba_dt_rank is _not_ the right field for how this is used
+        #   in llama.cpp
+        self.gguf_writer.add_ssm_time_step_rank(self.find_hparam(["n_heads"]))
+
+        ## Attention params ##
+        head_count_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
+        head_count_kv_vec = [
+            head_count_kv if i in self._attn_layers else 0 for i in range(self.block_count)
+        ]
+        if rope_dim := self.hparams.get("attn_rotary_emb"):
+            self.gguf_writer.add_rope_dimension_count(rope_dim)
+        self.gguf_writer.add_head_count_kv(head_count_kv_vec)
+
+        ## If Bamba, use rope, otherwise don't
+        use_rope = "BambaForCausalLM" in self.hparams["architectures"]
+        self.gguf_writer.add_rope_scaling_finetuned(use_rope)
+        if not use_rope:
+            self.gguf_writer.add_context_length(2**20)
+
+        ## Validation ##
+        d_head = self.find_hparam(["d_head"], optional=True) or 64
+        assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported"
+        assert self.d_inner % d_head == 0, f"SSM inner size {self.d_inner} not a multiple of head dim {d_head}"
+
+    def set_vocab(self):
+        self.hparams["pad_vocab_size_multiple"] = 8
+        Mamba2Model.set_vocab(self)
+
+
 @ModelBase.register("BailingMoeForCausalLM")
 class BailingMoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.BAILINGMOE
@@ -6567,7 +7281,7 @@ class FalconH1Model(Mamba2Model):
         # Use Llama conversion for attention
         self._transformer_model_class = LlamaModel
 
-        # n_group and d_inner are used during reshape_tensors for mamaba2
+        # n_group and d_inner are used during reshape_tensors for mamba2
         self.n_group = self.find_hparam(["n_groups"])
         self.d_inner = self.find_hparam(["mamba_d_ssm"])
         self.d_head = self.find_hparam(["d_head"])
@@ -6823,6 +7537,50 @@ class SmolLM3Model(LlamaModel):
             chat_template = tokenizer.chat_template.replace("[:]", "")
             self.gguf_writer.add_chat_template(chat_template)
 
+
+@ModelBase.register("Lfm2ForCausalLM")
+@ModelBase.register("LFM2ForCausalLM")
+class LFM2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.LFM2
+
+    def _add_feed_forward_length(self):
+        ff_dim = self.hparams["block_ff_dim"]
+
+        auto_adjust_ff_dim = self.hparams["block_auto_adjust_ff_dim"]
+        ff_dim = self.hparams["block_ff_dim"]
+        ffn_dim_multiplier = self.hparams["block_ffn_dim_multiplier"]
+        multiple_of = self.hparams["block_multiple_of"]
+
+        if auto_adjust_ff_dim:
+            ff_dim = int(2 * ff_dim / 3)
+            # custom dim factor multiplier
+            if ffn_dim_multiplier is not None:
+                ff_dim = int(ffn_dim_multiplier * ff_dim)
+            ff_dim = multiple_of * ((ff_dim + multiple_of - 1) // multiple_of)
+
+        self.gguf_writer.add_feed_forward_length(ff_dim)
+
+    def set_gguf_parameters(self):
+        # set num_key_value_heads only for attention layers
+        self.hparams["num_key_value_heads"] = [
+            self.hparams["num_key_value_heads"] if layer_type == "full_attention" else 0
+            for layer_type in self.hparams["layer_types"]
+        ]
+
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+        self.gguf_writer.add_shortconv_l_cache(self.hparams["conv_L_cache"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["norm_eps"])
+        self._add_feed_forward_length()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # conv op requires 2d tensor
+        if 'conv.conv' in name:
+            data_torch = data_torch.squeeze(1)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 ###### 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
@@ -128,6 +126,10 @@ models = [
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
+    {"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
@@ -143,11 +145,12 @@ pre_computed_hashes = [
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
+    {"name": "kimi-k2",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base",   "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
 ]
 
 
 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)
@@ -228,7 +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")
@@ -236,11 +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
@@ -250,15 +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/build.md

@@ -305,9 +305,8 @@ 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 +333,7 @@ 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 +356,8 @@ 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 +373,9 @@ 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 +385,29 @@ 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 +554,23 @@ 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/ops.md

@@ -0,0 +1,95 @@
+# GGML Operations
+
+List of GGML operations and backend support status.
+
+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 | ❌ | ✅ | ✅ | 🟡 |

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

@@ -107,7 +107,7 @@ 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;
 

ggml/CMakeLists.txt

@@ -181,6 +181,8 @@ 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 +272,7 @@ 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/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/include/ggml.h

@@ -1297,6 +1297,19 @@ extern "C" {
             struct ggml_tensor  * a,
             float                 s);
 
+    // x = s * a + b
+    GGML_API struct ggml_tensor * ggml_scale_bias(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
+    GGML_API struct ggml_tensor * ggml_scale_bias_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set(
             struct ggml_context * ctx,

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");
 

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

@@ -2090,6 +2090,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             {
                 // 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_CPY: {
@@ -2188,7 +2189,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_CLAMP:
@@ -2210,6 +2210,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_SCALE:
+            float bias;
+            memcpy(&bias, (float*)op->op_params + 1, sizeof(float));
+            return bias == 0.0f; // TODO: support bias != 0.0f
         case GGML_OP_SOFT_MAX:
             // TODO: support broadcast
             // ref: https://github.com/ggml-org/llama.cpp/pull/14435

ggml/src/ggml-cpu/CMakeLists.txt

@@ -494,9 +494,9 @@ 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/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);
             }
         }
@@ -4643,9 +4646,11 @@ static void ggml_compute_forward_scale_f32(
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    // scale factor
-    float v;
-    memcpy(&v, dst->op_params, sizeof(float));
+    float s; // scale factor
+    float b; // bias
+
+    memcpy(&s, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&b, (float *) dst->op_params + 1, sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -4664,12 +4669,22 @@ static void ggml_compute_forward_scale_f32(
 
     const size_t nb1 = dst->nb[1];
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        if (dst->data != src0->data) {
-            // src0 is same shape as dst => same indices
-            memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float));
+    if (b == 0.0f) {
+        for (int i1 = ir0; i1 < ir1; i1++) {
+            if (dst->data != src0->data) {
+                // src0 is same shape as dst => same indices
+                // TODO: add x parameter to ggml_vec_scale_f32 and remove this memcpy
+                memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float));
+            }
+            ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), s);
+        }
+    } else {
+        for (int i1 = ir0; i1 < ir1; i1++) {
+            ggml_vec_mad1_f32(nc,
+                (float *) ((char *) dst->data  + i1*nb1),
+                (float *) ((char *) src0->data + i1*nb1),
+                s, b);
         }
-        ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), v);
     }
 }
 

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-cpu/vec.h

@@ -351,6 +351,45 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
 #endif
 }
 
+inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, const float s, const float b) {
+#if defined(GGML_USE_ACCELERATE)
+    vDSP_vsmsa(x, 1, &s, &b, y, 1, n);
+#elif defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+        // scalar ; TODO: Write SVE code
+        for (int i = 0; i < n; ++i) {
+            y[i] = x[i]*s + b;
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
+
+        GGML_F32_VEC vs = GGML_F32_VEC_SET1(s);
+        GGML_F32_VEC vb = GGML_F32_VEC_SET1(b);
+
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ay[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_FMA(ay[j], vs, vb);
+
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] = x[i]*s + b;
+        }
+    #endif
+#else
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = x[i]*s + b;
+    }
+#endif
+}
+
 //inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) { for (int i = 0; i < n; ++i) y[i] *= v;          }
 inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #if defined(GGML_USE_ACCELERATE)

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/cpy-utils.cuh

@@ -0,0 +1,251 @@
+#pragma once
+
+#include "ggml-common.h"
+
+static __device__ __forceinline__ void convert_f32_f32(const float * src, float * dst) {
+    *dst = *src;
+}
+
+static __device__ __forceinline__ void convert_f32_f16(const float * src, half * dst) {
+    *dst = __float2half(*src);
+}
+
+static __device__ __forceinline__ void convert_f32_bf16(const float * src, nv_bfloat16 * dst) {
+    *dst = *src;
+}
+
+static __device__ __forceinline__ void convert_f16_f16(const half * src, half * dst) {
+    *dst = *src;
+}
+
+static __device__ __forceinline__ void convert_f16_f32(const half * src, float * dst) {
+    *dst = *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);
+}
+
+static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
+    convert_f32_f32((const float *)cxi, (float *)cdsti);
+}
+
+static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
+    convert_f32_f16((const float *)cxi, (half *)cdsti);
+}
+
+static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
+    convert_f32_bf16((const float *)cxi, (nv_bfloat16 *)cdsti);
+}
+
+static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
+    convert_f16_f16((const half *)cxi, (half *)cdsti);
+}
+
+static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
+    convert_f16_f32((const half *)cxi, (float *)cdsti);
+}

ggml/src/ggml-cuda/cpy.cu

@@ -1,46 +1,12 @@
 #include "cpy.cuh"
 #include "dequantize.cuh"
+#include "cpy-utils.cuh"
 #ifdef GGML_USE_MUSA
 #include "ggml-musa/mudnn.cuh"
 #endif // GGML_USE_MUSA
 
 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,
@@ -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,

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

@@ -33,8 +33,10 @@ typedef void (* fattn_kernel_t)(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -521,7 +523,7 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
 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 +537,8 @@ 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 +547,15 @@ 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 +574,7 @@ 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;
@@ -617,16 +620,31 @@ static __global__ void flash_attn_combine_results(
         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 +662,11 @@ 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 +723,6 @@ 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();
@@ -853,8 +869,8 @@ void launch_fattn(
         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,
+        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,
         Q->nb[1], Q->nb[2], Q->nb[3],
         nb11, nb12, nb13,
         nb21, nb22, nb23,
@@ -869,11 +885,11 @@ 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

@@ -1224,8 +1224,10 @@ static __global__ void flash_attn_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -1274,8 +1276,8 @@ 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).
@@ -1285,18 +1287,19 @@ 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;
@@ -1325,18 +1328,19 @@ 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;

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

@@ -31,8 +31,10 @@ static __global__ void flash_attn_tile_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -62,15 +64,17 @@ 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.");
@@ -255,6 +259,8 @@ 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 +272,21 @@ 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
@@ -290,8 +296,8 @@ 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/src/ggml-cuda/fattn-tile-f32.cu

@@ -31,8 +31,10 @@ static __global__ void flash_attn_tile_ext_f32(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -74,15 +76,17 @@ 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.");
 
@@ -265,6 +269,8 @@ 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;
@@ -276,22 +282,22 @@ 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/src/ggml-cuda/fattn-vec-f16.cuh

@@ -28,8 +28,10 @@ static __global__ void flash_attn_vec_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -65,14 +67,16 @@ 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.");
@@ -330,12 +334,11 @@ 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);
@@ -344,8 +347,8 @@ static __global__ void flash_attn_vec_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(ne32);
+    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/src/ggml-cuda/fattn-vec-f32.cuh

@@ -28,8 +28,10 @@ static __global__ void flash_attn_vec_ext_f32(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -53,8 +55,8 @@ 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);
@@ -77,14 +79,16 @@ 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;
@@ -326,12 +330,11 @@ 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);
@@ -340,8 +343,8 @@ 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/src/ggml-cuda/fattn-wmma-f16.cu

@@ -47,8 +47,10 @@ static __global__ void flash_attn_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -95,17 +97,19 @@ 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);
 
@@ -400,7 +404,6 @@ 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 +412,8 @@ 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 +424,7 @@ 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 +438,7 @@ 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);
@@ -442,7 +447,8 @@ 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);

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

@@ -43,6 +43,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set-rows.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -2230,6 +2231,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GET_ROWS_BACK:
             ggml_cuda_op_get_rows_back(ctx, dst);
             break;
+        case GGML_OP_SET_ROWS:
+            ggml_cuda_op_set_rows(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2299,6 +2303,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_EXP:
                     ggml_cuda_op_exp(ctx, dst);
                     break;
+                case GGML_UNARY_OP_ELU:
+                    ggml_cuda_op_elu(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2583,6 +2590,9 @@ 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];
 
@@ -2604,9 +2614,12 @@ 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]);
@@ -3112,6 +3125,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_ELU:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -3216,6 +3230,14 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                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;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -3335,8 +3357,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SSM_SCAN: {
             if (op->src[3]->ne[0] == 1) {
                 // Mamba2
-                // (kernel only supports d_state == 128 && d_head % 16 == 0)
-                return op->src[0]->ne[0] == 128 && op->src[0]->ne[1] % 16 == 0;
+                // (kernel only supports (d_state == 128 || d_state == 256) && d_head % 16 == 0)
+                return (op->src[0]->ne[0] == 128 || op->src[0]->ne[0] == 256) && op->src[0]->ne[1] % 16 == 0;
             } else {
                 // Mamba
                 // (kernel only supports d_state == 16, d_head == 1, n_head % 128 == 0, n_group == 1)
@@ -3398,12 +3420,6 @@ 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;
             }
@@ -3416,6 +3432,9 @@ 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/scale.cu

@@ -1,18 +1,18 @@
 #include "scale.cuh"
 
-static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) {
+static __global__ void scale_f32(const float * x, float * dst, const float scale, const float bias, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= k) {
         return;
     }
 
-    dst[i] = scale * x[i];
+    dst[i] = scale * x[i] + bias;
 }
 
-static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
+static void scale_f32_cuda(const float * x, float * dst, const float scale, const float bias, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
-    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
+    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, bias, k);
 }
 
 void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -25,7 +25,9 @@ void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     float scale;
-    memcpy(&scale, dst->op_params, sizeof(float));
+    float bias;
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&bias,  (float *) dst->op_params + 1, sizeof(float));
 
-    scale_f32_cuda(src0_d, dst_d, scale, ggml_nelements(src0), stream);
+    scale_f32_cuda(src0_d, dst_d, scale, bias, ggml_nelements(src0), stream);
 }

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

@@ -0,0 +1,288 @@
+#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);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, half>(const float * src_f, half * dst_h) {
+    convert_f32_f16(src_f, dst_h);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, nv_bfloat16>(const float * src_f, nv_bfloat16 * dst_b) {
+    convert_f32_bf16(src_f, dst_b);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, float>(const float * src_f, float * dst_f) {
+    convert_f32_f32(src_f, dst_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);
+}
+
+// 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) {
+
+    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>
+static __global__ void k_set_rows(
+        const src_t * __restrict__ src0, const int64_t * __restrict__ src1, dst_t * __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;
+
+    if (i >= ne_total) {
+        return;
+    }
+
+    const int64_t i03 = i / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i - 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 src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
+    dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
+
+    const src_t* src_elem = src0_row + i00;
+    dst_t* dst_elem = dst_row_ptr + i00;
+    set_rows_1(src_elem, dst_elem);
+
+    GGML_UNUSED(ne10);
+    GGML_UNUSED(ne13);
+}
+
+template<typename src_t, typename dst_t>
+static void set_rows_cuda(
+        const src_t * src0_d, const int64_t * src1_d, dst_t * 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) {
+
+    const int64_t ne_total = ne00 * ne01 * ne02 * ne03;
+    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(src_t);
+    const int64_t s02 = nb02/sizeof(src_t);
+    const int64_t s03 = nb03/sizeof(src_t);
+    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/sizeof(dst_t);
+    const int64_t s2  = nb2/sizeof(dst_t);
+    const int64_t s3  = nb3/sizeof(dst_t);
+
+    if (ne_total > 0) {
+        k_set_rows<<<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);
+    }
+}
+
+
+void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_I64);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const float * src0_d   = (const float *)src0->data;
+    const int64_t * src1_d = (const int64_t *)src1->data;
+
+    cudaStream_t stream = ctx.stream();
+
+
+
+    if (dst->type == GGML_TYPE_F32) {
+        set_rows_cuda(
+            src0_d, src1_d, (float*)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_F16) {
+        set_rows_cuda(
+            src0_d, src1_d, (half*)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_BF16) {
+        set_rows_cuda(
+            src0_d, src1_d, (nv_bfloat16*)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_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 %s", ggml_type_name(dst->type));
+    }
+}

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

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+#define CUDA_SET_ROWS_BLOCK_SIZE 256
+
+void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/ssm-conv.cu

@@ -107,8 +107,11 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
         if (nc == 4) {
             ssm_conv_f32<threads, 4><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
                                                                      dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+        } else if (nc == 3) {
+            ssm_conv_f32<threads, 3><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
+                                                                     dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
-            GGML_ABORT("Only support kernel size = 4  now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     } else {
         if (nc == 4) {
@@ -116,8 +119,13 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
             ssm_conv_long_token_f32<threads, 4, split_n_t><<<blocks, threads, 0, stream>>>(
                 src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+        } else if (nc == 3) {
+            const int64_t split_n_t = 32;
+            dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
+            ssm_conv_long_token_f32<threads, 3, split_n_t><<<blocks, threads, 0, stream>>>(
+                src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
-            GGML_ABORT("Only support kernel size = 4 right now.");
+            GGML_ABORT("Only support kernel size = 3 or size = 4 right now.");
         }
     }
 }

ggml/src/ggml-cuda/ssm-scan.cu

@@ -201,11 +201,11 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
                               const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
                               const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
                               cudaStream_t stream) {
-    const int threads = 128;
     // NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
     if (src3_nb1 == sizeof(float)) {
         // Mamba-2
         if (d_state == 128) {
+            const int threads = 128;
             GGML_ASSERT(d_state % threads == 0);
             // NOTE: can be any power of two between 4 and 64
             const int splitH = 16;
@@ -215,10 +215,21 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
                     src0, src1, src2, src3, src4, src5, src6, dst,
                     src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
                     src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
+        } else if (d_state == 256) { // Falcon-H1
+            const int threads = 256;
+            // NOTE: can be any power of two between 8 and 64
+            const int splitH = 16;
+            GGML_ASSERT(head_dim % splitH == 0);
+            const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
+            ssm_scan_f32_group<16, 256><<<blocks, threads, 0, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
         } else {
-            GGML_ABORT("doesn't support d_state!=128.");
+            GGML_ABORT("doesn't support d_state!=(128 or 256).");
         }
     } else {
+        const int threads = 128;
         // Mamba-1
         GGML_ASSERT(n_head % threads == 0);
         GGML_ASSERT(head_dim == 1);

ggml/src/ggml-cuda/unary.cu

@@ -83,6 +83,10 @@ static __device__ __forceinline__ float op_log(float x) {
     return logf(x);
 }
 
+static __device__ __forceinline__ float op_elu(float x) {
+    return (x > 0.f) ? x : expm1f(x);
+}
+
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -196,6 +200,9 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_log>(ctx, dst);
 }
 
+void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_elu>(ctx, dst);
+}
 /* gated ops */
 
 template <float (*op)(float), typename T>

ggml/src/ggml-cuda/unary.cuh

@@ -59,6 +59,8 @@ void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -10,9 +10,6 @@
 #include "rocblas/rocblas.h"
 #endif // __HIP_PLATFORM_AMD__
 
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N HIPBLAS_OP_N
@@ -30,7 +27,6 @@
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
 #define cublasCreate hipblasCreate
 #define cublasDestroy hipblasDestroy
 #define cublasGemmEx hipblasGemmEx
@@ -42,7 +38,6 @@
 #define cublasSgemm hipblasSgemm
 #define cublasStatus_t hipblasStatus_t
 #define cublasOperation_t hipblasOperation_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
 #define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
@@ -144,6 +139,20 @@
 #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
+#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
+#define cublasComputeType_t hipblasComputeType_t
+#define cudaDataType_t hipDataType
+#else
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define cublasComputeType_t hipblasDatatype_t
+#define cudaDataType_t hipblasDatatype_t
+#endif
+
 #define __CUDA_ARCH__ 1300
 
 #if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)

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 {

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,
@@ -173,6 +214,12 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SILU,
     GGML_METAL_KERNEL_TYPE_SILU_4,
     GGML_METAL_KERNEL_TYPE_ELU,
+    GGML_METAL_KERNEL_TYPE_ABS,
+    GGML_METAL_KERNEL_TYPE_SGN,
+    GGML_METAL_KERNEL_TYPE_STEP,
+    GGML_METAL_KERNEL_TYPE_HARDSWISH,
+    GGML_METAL_KERNEL_TYPE_HARDSIGMOID,
+    GGML_METAL_KERNEL_TYPE_EXP,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
@@ -212,6 +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,
@@ -1129,13 +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);
@@ -1155,6 +1218,12 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                            silu,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                          silu_4,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ELU,                             elu,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ABS,                             abs,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SGN,                             sgn,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_STEP,                            step,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_HARDSWISH,                       hardswish,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_HARDSIGMOID,                     hardsigmoid,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_EXP,                             exp,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,                    soft_max_f16,                    has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,                  soft_max_f16_4,                  has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,                    soft_max_f32,                    has_simdgroup_reduction);
@@ -1194,6 +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);
@@ -1688,6 +1759,12 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_NEG:
+                case GGML_UNARY_OP_ABS:
+                case GGML_UNARY_OP_SGN:
+                case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_EXP:
                     return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
@@ -1875,7 +1952,7 @@ 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,
           id<MTLComputeCommandEncoder>   encoder,
@@ -1885,7 +1962,10 @@ 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));
 
@@ -1895,7 +1975,7 @@ static bool ggml_metal_encode_node(
     struct ggml_tensor * dst  = node;
 
     if (ggml_is_empty(dst)) {
-        return true;
+        return 1;
     }
 
     switch (dst->op) {
@@ -1906,7 +1986,7 @@ static bool ggml_metal_encode_node(
         case GGML_OP_PERMUTE:
             {
                 // noop -> next node
-            } return true;
+            } return 1;
         default:
             {
             } break;
@@ -1973,6 +2053,8 @@ 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) {
@@ -2044,37 +2126,15 @@ 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,
@@ -2101,12 +2161,117 @@ 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;
+
+                    for (n_fuse = 0; n_fuse <= 6; ++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) {
@@ -2114,7 +2279,11 @@ 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)];
                 }
@@ -2239,12 +2408,13 @@ 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);
@@ -2256,7 +2426,9 @@ static bool ggml_metal_encode_node(
                 GGML_ASSERT(ggml_is_contiguous(src0));
 
                 float scale;
-                memcpy(&scale, dst->op_params, sizeof(scale));
+                float bias;
+                memcpy(&scale, ((const int32_t *) dst->op_params) + 0, sizeof(float));
+                memcpy(&bias,  ((const int32_t *) dst->op_params) + 1, sizeof(float));
 
                 int64_t n = ggml_nelements(dst);
 
@@ -2273,6 +2445,7 @@ static bool ggml_metal_encode_node(
                 [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
                 [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+                [encoder setBytes:&bias  length:sizeof(bias)  atIndex:3];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
             } break;
@@ -2436,6 +2609,78 @@ static bool ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } break;
+                case GGML_UNARY_OP_ABS:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ABS].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_SGN:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SGN].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_STEP:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_STEP].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_HARDSWISH:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_HARDSWISH].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_HARDSIGMOID:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_HARDSIGMOID].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_EXP:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_EXP].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                 default:
                 {
                     GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
@@ -2671,7 +2916,7 @@ 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;
@@ -3547,7 +3792,7 @@ 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;
@@ -3563,7 +3808,7 @@ 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
@@ -3571,7 +3816,7 @@ 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
@@ -3580,7 +3825,7 @@ 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;
                     }
 
                     {
@@ -4012,12 +4257,95 @@ 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; ++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
 
@@ -4028,23 +4356,16 @@ 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:
             {
@@ -5439,7 +5760,7 @@ static bool ggml_metal_encode_node(
             }
     }
 
-    return true;
+    return n_fuse;
 }
 
 static enum ggml_status ggml_metal_graph_compute(
@@ -5945,20 +6266,22 @@ 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, encoder, mem_pool);
 
             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;
@@ -970,60 +994,161 @@ 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,
         constant     float & scale,
+        constant     float & bias,
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * scale;
+    dst[tpig] = src0[tpig] * scale + bias;
 }
 
 kernel void kernel_scale_4(
         device const float4 * src0,
         device       float4 * dst,
         constant     float  & scale,
+        constant     float  & bias,
         uint tpig[[thread_position_in_grid]]) {
-    dst[tpig] = src0[tpig] * scale;
+    dst[tpig] = src0[tpig] * scale + bias;
 }
 
 kernel void kernel_clamp(
@@ -1197,6 +1322,51 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+kernel void kernel_abs(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = fabs(src0[tpig]);
+}
+
+kernel void kernel_sgn(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = (x > 0.0f) ? 1.0f : ((x < 0.0f) ? -1.0f : 0.0f);
+}
+
+kernel void kernel_step(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] > 0.0f ? 1.0f : 0.0f;
+}
+
+kernel void kernel_hardswish(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = x * fmin(1.0f, fmax(0.0f, (x + 3.0f) / 6.0f));
+}
+
+kernel void kernel_hardsigmoid(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = fmin(1.0f, fmax(0.0f, (x + 3.0f) / 6.0f));
+}
+
+kernel void kernel_exp(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = exp(src0[tpig]);
+}
+
 kernel void kernel_reglu(
         device const char * src0,
         device const char * src1,
@@ -2069,26 +2239,39 @@ 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);
@@ -2107,12 +2290,26 @@ 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-opencl/CMakeLists.txt

@@ -88,6 +88,7 @@ set(GGML_OPENCL_KERNELS
     rms_norm
     rope
     scale
+    set_rows
     sigmoid
     silu
     softmax_4_f32
@@ -103,6 +104,9 @@ set(GGML_OPENCL_KERNELS
     tanh
     pad
     repeat
+    mul_mat_f16_f32
+    conv2d
+    conv2d_f16_f32
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

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

@@ -351,6 +351,7 @@ struct ggml_backend_opencl_context {
     cl_program program_gemv_noshuffle_general;
     cl_program program_gemv_noshuffle;
     cl_program program_get_rows;
+    cl_program program_set_rows;
     cl_program program_glu;
     cl_program program_im2col_f16;
     cl_program program_im2col_f32;
@@ -367,6 +368,7 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mv_f16_f32;
     cl_program program_mul_mv_f32_f32;
     cl_program program_mul;
+    cl_program program_mul_mat_f16_f32_tiled;
     cl_program program_div;
     cl_program program_sub;
     cl_program program_norm;
@@ -388,6 +390,9 @@ struct ggml_backend_opencl_context {
     cl_program program_tanh;
     cl_program program_upscale;
     cl_program program_concat;
+    cl_program program_conv_2d_f16;
+    cl_program program_conv_2d_f32;
+    cl_program program_conv_2d_f16_f32;
     cl_program program_tsembd;
     cl_program program_mul_mv_id_q4_0_f32_8x_flat;
 
@@ -412,6 +417,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
+    cl_kernel kernel_set_rows_f32, kernel_set_rows_f16;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
     cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
     cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
@@ -420,6 +426,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mat_f16_f32_1row;
     cl_kernel kernel_mul_mat_f16_f32;
     cl_kernel kernel_mul_mat_f16_f32_l4;
+    cl_kernel kernel_mul_mat_f16_f32_tiled;
     cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
     cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
@@ -437,6 +444,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_upscale_bilinear;
     cl_kernel kernel_concat_f32_contiguous;
     cl_kernel kernel_concat_f32_non_contiguous;
+    cl_kernel kernel_conv_2d_f16;
+    cl_kernel kernel_conv_2d_f32;
+    cl_kernel kernel_conv_2d_f16_f32;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
 
@@ -529,6 +539,16 @@ struct ggml_backend_opencl_context {
         fclose(ftrace);
     }
 
+    size_t get_kernel_workgroup_size(cl_kernel kernel) const {
+        size_t workgroup_size = 0;
+        size_t ret_size = 0;
+        CL_CHECK(
+            clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE,
+                sizeof(size_t), &workgroup_size, &ret_size));
+        GGML_ASSERT(sizeof(size_t) == ret_size);
+        return workgroup_size;
+    }
+
     void enqueue_ndrange_kernel(cl_kernel kernel, cl_uint work_dim, size_t *global_work_size, size_t *local_work_size, const ggml_tensor * tensor) {
 #ifdef GGML_OPENCL_PROFILING
         cl_event evt;
@@ -1003,6 +1023,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mat_f16_f32_tiled
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mat_f16_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mat_f16_f32.cl");
+#endif
+        backend_ctx->program_mul_mat_f16_f32_tiled =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f32_tiled = clCreateKernel(backend_ctx->program_mul_mat_f16_f32_tiled, "mul_mat_f16_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1431,6 +1467,64 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         }
     }
 
+    // set_rows
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "set_rows.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("set_rows.cl");
+#endif
+        backend_ctx->program_set_rows =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_set_rows_f32  = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16  = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+     // conv2d
+     {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src {
+                    #include "conv2d.cl.h"
+                };
+                const std::string kernel_src_f16_f32 {
+                    #include "conv2d_f16_f32.cl.h"
+                };
+        #else
+                const std::string kernel_src = read_file("conv2d.cl");
+                const std::string kernel_src_f16_f32 = read_file("conv2d_f16_f32.cl");
+        #endif
+                if (!kernel_src.empty()) {
+                    backend_ctx->program_conv_2d_f16 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), (std::string(compile_opts) + " -DUSE_FP16=1").c_str());
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f16 = clCreateKernel(backend_ctx->program_conv_2d_f16, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                    backend_ctx->program_conv_2d_f32 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f32 = clCreateKernel(backend_ctx->program_conv_2d_f32, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                } else {
+                    GGML_LOG_WARN("ggml_opencl: conv2d kernel source not found or empty. This op will not be available.\n");
+                    backend_ctx->program_conv_2d_f16 = nullptr;
+                    backend_ctx->kernel_conv_2d_f16 = nullptr;
+                    backend_ctx->program_conv_2d_f32 = nullptr;
+                    backend_ctx->kernel_conv_2d_f32 = nullptr;
+                }
+                if (!kernel_src_f16_f32.empty()) {
+                    backend_ctx->program_conv_2d_f16_f32 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16_f32.c_str(), compile_opts);
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f16_f32 = clCreateKernel(backend_ctx->program_conv_2d_f16_f32, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                } else {
+                    GGML_LOG_WARN("ggml_opencl: conv2d_f16_f32 kernel source not found or empty. This op will not be available.\n");
+                    backend_ctx->program_conv_2d_f16_f32 = nullptr;
+                    backend_ctx->kernel_conv_2d_f16_f32 = nullptr;
+                }
+    }
+
     // mul_mv_id_q4_0_f32_8x_flat
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2233,8 +2327,18 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             {
                 // TODO: add support
                 // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-                return false;
-            } break;
+#pragma message("TODO: implement BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                if (op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                switch (op->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                        return true;
+                    default:
+                        return false;
+                }
+            }
         case GGML_OP_CPY:
         case GGML_OP_DUP:
         case GGML_OP_CONT:
@@ -2304,6 +2408,10 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                    op->src[0]->ne[3] == 1 && op->ne[3] == 1;
         case GGML_OP_UPSCALE:
             return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_CONV_2D:
+            return (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16) ||
+                   (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                   (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
         case GGML_OP_CONCAT:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -3374,6 +3482,111 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
+static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    // ne0 = ne00
+    // ne2 = ne02
+    // ne3 = ne03
+
+    const int      ne01 = src0->ne[1];
+    const int      ne02 = src0->ne[2];
+    const int      ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int      ne11 = src1->ne[1];
+    const int      ne12 = src1->ne[2];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int      ne0  = dst->ne[0];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    const int nblk0 = ne0/ggml_blck_size(dst->type);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            kernel = backend_ctx->kernel_set_rows_f32;
+            break;
+        case GGML_TYPE_F16:
+            kernel = backend_ctx->kernel_set_rows_f16;
+            break;
+        default:
+            GGML_ABORT("not implemented");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &nblk0));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb3));
+
+    int nth0 = 64;
+    if (backend_ctx->gpu_family == INTEL) {
+        nth0 = 32;
+    } else if (backend_ctx->gpu_family == ADRENO) {
+        nth0 = 64;
+    }
+
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth0 < nblk0 && nth0 < max_workgroup_size) {
+        nth0 *= 2;
+    }
+
+    int rows_per_workgroup = 1;
+    if (nth0 > nblk0) {
+        rows_per_workgroup = nth0 / nblk0;
+        nth0 = nblk0;
+    }
+
+    size_t global_work_size[] = {
+        (size_t)(ne01 + rows_per_workgroup - 1)/rows_per_workgroup*nth0,
+        (size_t)ne02*rows_per_workgroup,
+        (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth0, (size_t)rows_per_workgroup, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4784,6 +4997,135 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
 }
 
+static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int M = src0->ne[1];
+    const int N = src1->ne[1];
+    const int K = src0->ne[0];
+
+    cl_kernel kernel = backend_ctx->kernel_mul_mat_f16_f32_tiled;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(int),      &M));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(int),      &N));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),      &K));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &offsetd));
+
+    // Tiling parameters. These need to be tuned for optimal performance.
+    // They must match the #defines in the kernel mul_mat_f16_f32.cl.
+    //
+    // OPWM / OPWN: Output tile size per Work-Group. A work-group computes a tile of size OPWM x OPWN.
+    // TPWM / TPWN: Threads per Work-group. This is the work-group size.
+    // OPTM / OPTN: Output elements per Thread. Each thread computes OPTM x OPTN elements.
+    //
+    // The following relationships must hold:
+    //   OPWM = TPWM * OPTM
+    //   OPWN = TPWN * OPTN
+    //
+    const int OPWM = 64;
+    const int OPWN = 64;
+    const int TPWM = 16;
+    const int TPWN = 8;
+
+    size_t local_work_size[2] = { TPWM, TPWN };
+    size_t global_work_size[2] = {
+        (size_t) ((M + OPWM - 1) / OPWM) * TPWM,
+        (size_t) ((N + OPWN - 1) / OPWN) * TPWN,
+    };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_conv_2d(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_TENSOR_BINARY_OP_LOCALS;
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const cl_uint Cout = ne03; const cl_uint Cin = ne02; const cl_uint N = ne13;
+    const cl_uint KW = ne00; const cl_uint KH = ne01; const cl_uint W = ne10; const cl_uint H = ne11; const cl_uint OW = ne0; const cl_uint OH = ne1;
+
+    const cl_uint s0 = dst->op_params[0]; const cl_uint s1 = dst->op_params[1];
+    const cl_uint p0 = dst->op_params[2]; const cl_uint p1 = dst->op_params[3];
+    const cl_uint d0 = dst->op_params[4]; const cl_uint d1 = dst->op_params[5];
+
+    const cl_uint cl_nb01 = nb01/ggml_type_size(src0->type); const cl_uint cl_nb02 = nb02/ggml_type_size(src0->type); const cl_uint cl_nb03 = nb03/ggml_type_size(src0->type);
+    const cl_uint cl_nb11 = nb11/ggml_type_size(src1->type); const cl_uint cl_nb12 = nb12/ggml_type_size(src1->type); const cl_uint cl_nb13 = nb13/ggml_type_size(src1->type);
+    const cl_uint cl_nb1 = nb1/ggml_type_size(dst->type); const cl_uint cl_nb2 = nb2/ggml_type_size(dst->type); const cl_uint cl_nb3 = nb3/ggml_type_size(dst->type);
+
+    const int64_t NPQ = (int64_t)N * OW * OH;
+
+    const uint32_t BS_K = 64;
+    const uint32_t BS_NPQ = 64;
+    const uint32_t BS_CRS = 16;
+    const uint32_t VEC_SIZE = 4;
+
+    const uint32_t TS_K = 4;
+    const uint32_t TS_NPQ = 8;
+
+    const uint32_t WG_K = BS_K / TS_K;
+    const uint32_t WG_NPQ = BS_NPQ / TS_NPQ;
+
+    auto splitWork = [](uint32_t work_size, uint32_t block_size) { return (block_size + work_size - 1) / block_size; };
+    const uint32_t NB_K = splitWork(Cout, BS_K);
+    const uint32_t NB_NPQ = splitWork(NPQ, BS_NPQ);
+
+    cl_kernel kernel;
+    size_t shmem_size;
+
+    if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_conv_2d_f16;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_half) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_half4));
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_conv_2d_f32;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_float) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_float4));
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_conv_2d_f16_f32;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_half) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_float4));
+    } else {
+        GGML_ASSERT(false && "Unsupported data type combination for conv2d");
+        return;
+    }
+
+    cl_uint idx = 0;
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extra0->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extra1->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extrad->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, idx++, shmem_size, NULL));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &Cout)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &Cin)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &N));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &KW)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &KH)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &W)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &H));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &OW)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &OH));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &s0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &s1)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &p0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &p1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &d0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &d1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb01)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb02)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb03));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb11)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb12)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb13));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb1)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb2)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb3));
+
+    size_t global_work_size[] = { (size_t)NB_K * WG_K, (size_t)NB_NPQ * WG_NPQ, 1 };
+    size_t local_work_size[] = { (size_t)WG_K, (size_t)WG_NPQ, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4797,6 +5139,18 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
+     if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
+        src0->ne[1] > 32 &&   // M > 32
+        src1->ne[1] > 32 &&   // N > 32
+        src0->ne[0] > 32 &&   // K > 32
+        src0->ne[2] == 1 && src0->ne[3] == 1 &&
+        src1->ne[2] == 1 && src1->ne[3] == 1 &&
+        ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+        backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
+        ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
+        return;
+    }
+
     ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
@@ -5587,7 +5941,9 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
     float scale;
-    memcpy(&scale, dst->op_params, sizeof(scale));
+    float bias;
+    memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(float));
+    memcpy(&bias,  ((int32_t *) dst->op_params) + 1, sizeof(float));
 
     ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
@@ -5602,6 +5958,7 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
     CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
     CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
     CL_CHECK(clSetKernelArg(kernel, 4, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(float),    &bias));
 
     int n = ggml_nelements(dst)/4;
 
@@ -6385,6 +6742,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_get_rows;
             break;
+        case GGML_OP_SET_ROWS:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_set_rows;
+            break;
         case GGML_OP_CPY:
             if (!any_on_device) {
                 return false;
@@ -6517,6 +6880,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             ggml_cl_upscale(backend, tensor->src[0], tensor);
             return true;
+        case GGML_OP_CONV_2D:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_conv_2d;
+            break;
         case GGML_OP_CONCAT:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/conv2d.cl

@@ -0,0 +1,185 @@
+#ifdef USE_FP16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#define T_FLOAT half
+#define T_FLOAT4 half4
+#define VSTORE_T_FLOAT4(data, offset, p) vstore_half4_rte(data, offset, p)
+#else
+#define T_FLOAT float
+#define T_FLOAT4 float4
+#define VSTORE_T_FLOAT4(data, offset, p) vstore4(data, offset, p)
+#endif
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define T_ACCUM float4
+#define VEC_SIZE 4
+
+#define BS_K 64
+#define BS_NPQ 64
+#define BS_CRS 16
+
+#define TS_K 4
+#define TS_NPQ 8
+
+#define WG_K (BS_K / TS_K)
+#define WG_NPQ (BS_NPQ / TS_NPQ)
+
+#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE)
+#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE)
+
+static inline uint splitWork(uint work_size, uint block_size){
+    return (work_size + block_size - 1) / block_size;
+}
+
+REQD_SUBGROUP_SIZE_128
+kernel void kernel_conv_2d(
+    global void* p_knl,
+    ulong off_knl,
+    global void* p_src,
+    ulong off_src,
+    global void* p_dst,
+    ulong off_dst,
+    local void* shared,
+    uint Cout, uint Cin, uint N,
+    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
+    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
+    uint nb01, uint nb02, uint nb03,
+    uint nb11, uint nb12, uint nb13,
+    uint nb1, uint nb2, uint nb3
+) {
+    global T_FLOAT* knl_data = (global T_FLOAT*) ((global char*)p_knl + off_knl);
+    global T_FLOAT* src_data = (global T_FLOAT*) ((global char*)p_src + off_src);
+    global T_FLOAT* dst_data = (global T_FLOAT*) ((global char*)p_dst + off_dst);
+
+    const uint K = Cout;
+    const uint CRS = Cin*KH*KW;
+    const uint NPQ = N*OH*OW;
+
+    const uint lid_k = get_local_id(0);
+    const uint lid_npq = get_local_id(1);
+    const uint tid = lid_npq * WG_K + lid_k;
+
+    const uint B_idx_K = get_group_id(0);
+    const uint B_idx_NPQ = get_group_id(1);
+
+    const uint offset_k = B_idx_K * BS_K;
+    const uint offset_npq = B_idx_NPQ * BS_NPQ;
+
+    local T_FLOAT* Ash = (local T_FLOAT*)shared;
+    local T_FLOAT4* Bsh = (local T_FLOAT4*) &Ash[BS_K * BS_CRS];
+
+    T_ACCUM regC[TS_K][TS_NPQ_VEC];
+    for (int i = 0; i < TS_K; ++i) {
+        for (int j = 0; j < TS_NPQ_VEC; ++j) {
+            regC[i][j] = (T_ACCUM)(0.0f);
+        }
+    }
+
+    const uint NB_CRS = splitWork(CRS, BS_CRS);
+
+    for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) {
+        const uint offset_crs = B_idx_CRS * BS_CRS;
+
+        for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) {
+            const uint k_l = i / BS_CRS;
+            const uint crs_l = i % BS_CRS;
+            const uint k_g = offset_k + k_l;
+            const uint crs_g = offset_crs + crs_l;
+
+            if (k_g < K && crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW*KH);
+                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
+                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
+            } else {
+                Ash[k_l * BS_CRS + crs_l] = (T_FLOAT)0.0f;
+            }
+        }
+
+        for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) {
+            const uint crs_l = i / BS_NPQ_VEC;
+            const uint npq_l_vec = i % BS_NPQ_VEC;
+            const uint crs_g = offset_crs + crs_l;
+
+            T_FLOAT4 val = (T_FLOAT4)(0.0f);
+            if (crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW * KH);
+                const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW;
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx = npq_g / (OH * OW);
+                        const uint pq_idx = npq_g % (OH * OW);
+                        const uint OH_idx = pq_idx / OW;
+                        const uint OW_idx = pq_idx % OW;
+                        const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1);
+                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);
+
+                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
+                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            ((T_FLOAT*)&val)[v] = src_data[src_idx];
+                        }
+                    }
+                }
+            }
+            Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) {
+            T_FLOAT regA[TS_K];
+            for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l];
+            }
+
+            for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+                T_FLOAT4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg];
+                for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                    regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), convert_float4(regB), regC[k_l_reg][npq_l_vec_reg]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+        const uint k_g = offset_k + lid_k * TS_K + k_l_reg;
+        if (k_g >= K) continue;
+
+        for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+            const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE;
+
+            const uint N_idx = npq_g_base / (OH * OW);
+            const uint pq_idx = npq_g_base % (OH * OW);
+            const uint OH_idx = pq_idx / OW;
+            const uint OW_idx = pq_idx % OW;
+
+            if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) {
+                const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3;
+                VSTORE_T_FLOAT4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]);
+            } else {
+                T_ACCUM res = regC[k_l_reg][npq_l_vec_reg];
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = npq_g_base + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx_s = npq_g / (OH*OW);
+                        const uint pq_idx_s = npq_g % (OH*OW);
+                        const uint OH_idx_s = pq_idx_s / OW;
+                        const uint OW_idx_s = pq_idx_s % OW;
+                        const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3;
+                        dst_data[dst_idx_s] = (T_FLOAT)(((float*)&res)[v]);
+                    }
+                }
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl

@@ -0,0 +1,176 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define T_ACCUM float4
+#define VEC_SIZE 4
+
+#define BS_K 64
+#define BS_NPQ 64
+#define BS_CRS 16
+
+#define TS_K 4
+#define TS_NPQ 8
+
+#define WG_K (BS_K / TS_K)
+#define WG_NPQ (BS_NPQ / TS_NPQ)
+
+#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE)
+#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE)
+
+static inline uint splitWork(uint work_size, uint block_size){
+    return (work_size + block_size - 1) / block_size;
+}
+
+REQD_SUBGROUP_SIZE_128
+kernel void kernel_conv_2d(
+    global void* p_knl,
+    ulong off_knl,
+    global void* p_src,
+    ulong off_src,
+    global void* p_dst,
+    ulong off_dst,
+    local void* shared,
+    uint Cout, uint Cin, uint N,
+    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
+    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
+    uint nb01, uint nb02, uint nb03,
+    uint nb11, uint nb12, uint nb13,
+    uint nb1, uint nb2, uint nb3
+) {
+    global half* knl_data = (global half*) ((global char*)p_knl + off_knl);
+    global float* src_data = (global float*) ((global char*)p_src + off_src);
+    global float* dst_data = (global float*) ((global char*)p_dst + off_dst);
+
+    const uint K = Cout;
+    const uint CRS = Cin*KH*KW;
+    const uint NPQ = N*OH*OW;
+
+    const uint lid_k = get_local_id(0);
+    const uint lid_npq = get_local_id(1);
+    const uint tid = lid_npq * WG_K + lid_k;
+
+    const uint B_idx_K = get_group_id(0);
+    const uint B_idx_NPQ = get_group_id(1);
+
+    const uint offset_k = B_idx_K * BS_K;
+    const uint offset_npq = B_idx_NPQ * BS_NPQ;
+
+    local half* Ash = (local half*)shared;
+    local float4* Bsh = (local float4*) &Ash[BS_K * BS_CRS];
+
+    T_ACCUM regC[TS_K][TS_NPQ_VEC];
+    for (int i = 0; i < TS_K; ++i) {
+        for (int j = 0; j < TS_NPQ_VEC; ++j) {
+            regC[i][j] = (T_ACCUM)(0.0f);
+        }
+    }
+
+    const uint NB_CRS = splitWork(CRS, BS_CRS);
+
+    for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) {
+        const uint offset_crs = B_idx_CRS * BS_CRS;
+
+        for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) {
+            const uint k_l = i / BS_CRS;
+            const uint crs_l = i % BS_CRS;
+            const uint k_g = offset_k + k_l;
+            const uint crs_g = offset_crs + crs_l;
+
+            if (k_g < K && crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW*KH);
+                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
+                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
+            } else {
+                Ash[k_l * BS_CRS + crs_l] = (half)0.0f;
+            }
+        }
+
+        for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) {
+            const uint crs_l = i / BS_NPQ_VEC;
+            const uint npq_l_vec = i % BS_NPQ_VEC;
+            const uint crs_g = offset_crs + crs_l;
+
+            float4 val = (float4)(0.0f);
+            if (crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW * KH);
+                const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW;
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx = npq_g / (OH * OW);
+                        const uint pq_idx = npq_g % (OH * OW);
+                        const uint OH_idx = pq_idx / OW;
+                        const uint OW_idx = pq_idx % OW;
+                        const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1);
+                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);
+
+                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
+                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            ((float*)&val)[v] = src_data[src_idx];
+                        }
+                    }
+                }
+            }
+            Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) {
+            half regA[TS_K];
+            for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l];
+            }
+
+            for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+                float4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg];
+                for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                    regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), regB, regC[k_l_reg][npq_l_vec_reg]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+        const uint k_g = offset_k + lid_k * TS_K + k_l_reg;
+        if (k_g >= K) continue;
+
+        for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+            const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE;
+
+            const uint N_idx = npq_g_base / (OH * OW);
+            const uint pq_idx = npq_g_base % (OH * OW);
+            const uint OH_idx = pq_idx / OW;
+            const uint OW_idx = pq_idx % OW;
+
+            if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) {
+                const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3;
+                vstore4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]);
+            } else {
+                T_ACCUM res = regC[k_l_reg][npq_l_vec_reg];
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = npq_g_base + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx_s = npq_g / (OH*OW);
+                        const uint pq_idx_s = npq_g % (OH*OW);
+                        const uint OH_idx_s = pq_idx_s / OW;
+                        const uint OW_idx_s = pq_idx_s % OW;
+                        const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3;
+                        dst_data[dst_idx_s] = ((float*)&res)[v];
+                    }
+                }
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/im2col_f16.cl

@@ -31,7 +31,7 @@ kernel void kernel_im2col_f16(
     src1 = (global float*)((global char*)src1 + offset1);
     dst = (global half*)((global char*)dst + offsetd);
 
-    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  ksize = OW * KH;
     long  kx = i / ksize;
     long  kd = kx * ksize;
     long  ky = (i - kd) / OW;

ggml/src/ggml-opencl/kernels/im2col_f32.cl

@@ -31,7 +31,7 @@ kernel void kernel_im2col_f32(
     src1 = (global float*)((global char*)src1 + offset1);
     dst = (global float*)((global char*)dst + offsetd);
 
-    long  ksize = OW * (KH > 1 ? KW : 1);
+    long  ksize = OW * KH;
     long  kx = i / ksize;
     long  kd = kx * ksize;
     long  ky = (i - kd) / OW;

ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl

@@ -0,0 +1,130 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define OPWM 64
+#define OPWN 64
+#define CPWK 8
+#define OPTM 4
+#define OPTN 8
+
+#define WG_M (OPWM / OPTM)
+#define WG_N (OPWN / OPTN)
+#define VEC_K (CPWK / 4)
+
+REQD_SUBGROUP_SIZE_128
+__kernel void mul_mat_f16_f32(
+    const int M, const int N, const int K,
+    __global const void* A_void, ulong A_offset,
+    __global const void* B_void, ulong B_offset,
+    __global       void* C_void, ulong C_offset) {
+
+    __global const half*  A = (__global const half* )((__global const char*)A_void + A_offset);
+    __global const float* B = (__global const float*)((__global const char*)B_void + B_offset);
+    __global       float* C = (__global       float*)((__global       char*)C_void + C_offset);
+
+    const int lidm = get_local_id(0);
+    const int lidn = get_local_id(1);
+    const int lid = lidn * WG_M + lidm;
+
+    const int offsetM = get_group_id(0) * OPWM;
+    const int offsetN = get_group_id(1) * OPWN;
+
+    __local half4  Alocal[OPWM][VEC_K];
+    __local float4 Blocal[OPWN][VEC_K];
+
+    float sum[OPTM][OPTN];
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        for (int wn = 0; wn < OPTN; wn++) {
+            sum[wm][wn] = 0.0f;
+        }
+    }
+
+    const int numTiles = (K + CPWK - 1) / CPWK;
+
+    const int load_row_a = lid % OPWM;
+    const int load_vec_k_a = lid / OPWM;
+    const int global_row_a = offsetM + load_row_a;
+
+    const int load_row_b = lid % OPWN;
+    const int load_vec_k_b = lid / OPWN;
+    const int global_row_b = offsetN + load_row_b;
+
+    for (int t = 0; t < numTiles; t++) {
+        const int k_start = t * CPWK;
+        const int k_vec_start_a = k_start + load_vec_k_a * 4;
+        const int k_vec_start_b = k_start + load_vec_k_b * 4;
+
+        if (global_row_a < M && k_vec_start_a < K) {
+            if (k_vec_start_a + 3 < K) {
+                Alocal[load_row_a][load_vec_k_a] = vload4(0, A + global_row_a * K + k_vec_start_a);
+            } else {
+                half4 tempA = (half4)(0.0h);
+                if (k_vec_start_a < K) tempA.s0 = A[global_row_a * K + k_vec_start_a];
+                if (k_vec_start_a + 1 < K) tempA.s1 = A[global_row_a * K + k_vec_start_a + 1];
+                if (k_vec_start_a + 2 < K) tempA.s2 = A[global_row_a * K + k_vec_start_a + 2];
+                Alocal[load_row_a][load_vec_k_a] = tempA;
+            }
+        } else {
+            Alocal[load_row_a][load_vec_k_a] = (half4)(0.0h);
+        }
+
+        if (global_row_b < N && k_vec_start_b < K) {
+            if (k_vec_start_b + 3 < K) {
+                Blocal[load_row_b][load_vec_k_b] = vload4(0, B + global_row_b * K + k_vec_start_b);
+            } else {
+                float4 tempB = (float4)(0.0f);
+                if (k_vec_start_b < K) tempB.s0 = B[global_row_b * K + k_vec_start_b];
+                if (k_vec_start_b + 1 < K) tempB.s1 = B[global_row_b * K + k_vec_start_b + 1];
+                if (k_vec_start_b + 2 < K) tempB.s2 = B[global_row_b * K + k_vec_start_b + 2];
+                Blocal[load_row_b][load_vec_k_b] = tempB;
+            }
+        } else {
+            Blocal[load_row_b][load_vec_k_b] = (float4)(0.0f);
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (int k_vec = 0; k_vec < VEC_K; k_vec++) {
+            float4 a_fvecs[OPTM];
+            int current_row_a = lidm;
+            for (int wm = 0; wm < OPTM; wm++) {
+                a_fvecs[wm] = convert_float4(Alocal[current_row_a][k_vec]);
+                current_row_a += WG_M;
+            }
+
+            float4 b_fvecs[OPTN];
+            int current_row_b = lidn;
+            for (int wn = 0; wn < OPTN; wn++) {
+                b_fvecs[wn] = Blocal[current_row_b][k_vec];
+                current_row_b += WG_N;
+            }
+
+            for (int wm = 0; wm < OPTM; wm++) {
+                for (int wn = 0; wn < OPTN; wn++) {
+                    sum[wm][wn] += dot(a_fvecs[wm], b_fvecs[wn]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        int globalRow = offsetM + lidm + wm * WG_M;
+        if (globalRow < M) {
+            for (int wn = 0; wn < OPTN; wn++) {
+                int globalCol = offsetN + lidn + wn * WG_N;
+                if (globalCol < N) {
+                    C[globalCol * M + globalRow] = sum[wm][wn];
+                }
+            }
+        }
+    }
+}

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

@@ -8,9 +8,10 @@ kernel void kernel_scale(
         ulong offset0,
         global float4 * dst,
         ulong offsetd,
-        float scale
+        float scale,
+        float bias
 ) {
     src0 = (global float4*)((global char*)src0 + offset0);
     dst = (global float4*)((global char*)dst + offsetd);
-    dst[get_global_id(0)] = src0[get_global_id(0)] * scale;
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
 }

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

@@ -0,0 +1,95 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_set_rows_f32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        int           ne11,
+        int           ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = i03%ne12;
+    int i11 = i02%ne11;
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global float * dst_row = (global float *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = (float)src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f16(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        int           ne11,
+        int           ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = i03%ne12;
+    int i11 = i02%ne11;
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global half  * dst_row = (global half  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
+    }
+}

ggml/src/ggml-sycl/backend.hpp

@@ -30,6 +30,7 @@
 #include "outprod.hpp"
 #include "quants.hpp"
 #include "rope.hpp"
+#include "set_rows.hpp"
 #include "softmax.hpp"
 #include "tsembd.hpp"
 #include "wkv.hpp"

ggml/src/ggml-sycl/gemm.hpp

@@ -32,39 +32,28 @@ public:
         else static_assert(0);
     }
 
-    // matrix A has m rows, k columns
-    // matrix B has k rows, n columns
-    // nra - number of elements to skip when moving into next row in A
-    // nrb - number of elements to skip when moving into next row in B
-    // nca - number of elements to skip when moving into next column in A
-    // ncb - number of elements to skip when moving into next column in B
-    // stride_a - number of elements to skip when moving to next A matrix
-    // stride_b - number of elements to skip when moving to next B matrix
-    // batches_a - number of A matrices
-    // batches_b - number of B matrices
     static void gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
-        const void * a, dt at, dnnl_dim_t nra, dnnl_dim_t nca, dnnl_dim_t stride_a,
-        const void * b, dt bt, dnnl_dim_t nrb, dnnl_dim_t ncb, dnnl_dim_t stride_b,
+        const void * a, dt at, dnnl_dim_t stra0, dnnl_dim_t stra1, dnnl_dim_t stra2,
+        const void * b, dt bt, dnnl_dim_t strb0, dnnl_dim_t strb1, dnnl_dim_t strb2,
         void * c, dt ct, const queue_ptr & q, dnnl_dim_t batches_a, dnnl_dim_t batches_b) {
 
         auto stream = ctx.stream_dnnl(q);
         auto eng = ctx.engine_dnnl(q);
 
-        // { # strides, # rows, # columns }
-        dnnl::memory::dims a_dims = { batches_a, m, k };
-        dnnl::memory::dims b_dims = { batches_b, k, n };
-        dnnl::memory::dims c_dims = { std::max(batches_a, batches_b), m, n };
-
-        // { # elements to skip to next stride, # elements to skip to next row, # elements to skip to next column }
-        dnnl::memory::dims a_strides = { stride_a, nra, nca };
-        dnnl::memory::dims b_strides = { stride_b, nrb, ncb };
-
+        dnnl::memory::dims a_dims = {batches_a, m, k };
+        dnnl::memory::dims a_strides = {stra2, stra1, stra0};
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_strides);
-        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_strides);
-        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::abc);
 
+        dnnl::memory::dims b_dims = {batches_b, k, n };
+        dnnl::memory::dims b_strides = {strb2, strb0, strb1};
+        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_strides);
+
+        dnnl::memory::dims c_dims = { std::max(batches_a, batches_b), m, n};
+        dnnl::memory::dims c_strides = {m*n, 1,  m };
+        const auto c_md    = dnnl::memory::desc(c_dims, ct, c_strides);
         dnnl::primitive_attr primitive_attr;
         primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+
 #ifdef GGML_SYCL_F16
         primitive_attr.set_fpmath_mode(dnnl::fpmath_mode::f16);
 #endif
@@ -76,24 +65,23 @@ public:
 
         auto scratchpad_md = matmul_pd.scratchpad_desc();
         auto scratchpad_mem = ctx.get_scratchpad_mem(scratchpad_md, eng, q);
+
         auto matmul_prim = dnnl::matmul(matmul_pd);
 
         std::unordered_map<int, dnnl::memory> matmul_args;
         matmul_args.insert({ DNNL_ARG_SRC, a_mem });
         matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
+
         matmul_args.insert({ DNNL_ARG_DST, c_mem });
         matmul_args.insert({ DNNL_ARG_SCRATCHPAD, scratchpad_mem });
 
         matmul_prim.execute(stream, matmul_args);
     }
 
-    // matrices A and B are column major, both having k rows
-    // matrix A has m column, matrix B has n columns
-    // output: column major matrix C = A transposed * B
     static void row_gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
         const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
 
-        gemm(ctx, m, n, k, a, at, k, 1, k * m, b, bt, 1, k, n * k, c, ct, q, 1, 1);
+        gemm(ctx, m, n, k, a, at, 1, k, k * m, b, bt, 1, k, n * k, c, ct, q, 1, 1);
     }
 };
 

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

@@ -41,6 +41,7 @@
 #include "ggml-sycl/element_wise.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/gemm.hpp"
+#include "ggml-sycl/set_rows.hpp"
 #include "ggml-sycl/sycl_hw.hpp"
 #include "ggml-sycl/getrows.hpp"
 #include "ggml.h"
@@ -1545,7 +1546,7 @@ static void mul_mat_p021_f16_f32(
 
 static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor,
+    const int row_stride_x, const int channel_stride_x,const int channel_stride_y, const int channel_x_divisor,
     const sycl::nd_item<3> &item_ct1) {
 
     const sycl::half *x = (const sycl::half *)vx;
@@ -1556,7 +1557,6 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
                         item_ct1.get_local_id(0);
     const int channel_x = channel / channel_x_divisor;
 
-    const int nrows_y   = ncols_x;
     const int nrows_dst = nrows_x;
     const int row_dst   = row_x;
 
@@ -1575,7 +1575,7 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
         const int row_y = col_x;
 
         const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
+        const int iy = channel * channel_stride_y + row_y;
 
         const float xi =
             sycl::vec<sycl::half, 1>(x[ix])
@@ -1695,7 +1695,7 @@ static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, con
     dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
 }
 
-static void scale_f32(const float * x, float * dst, const float scale, const int k,
+static void scale_f32(const float * x, float * dst, const float scale, const float bias, const int k,
                       const sycl::nd_item<3> &item_ct1) {
     const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
                   item_ct1.get_local_id(2);
@@ -1704,7 +1704,7 @@ static void scale_f32(const float * x, float * dst, const float scale, const int
         return;
     }
 
-    dst[i] = scale * x[i];
+    dst[i] = scale * x[i] + bias;
 }
 
 
@@ -1822,7 +1822,7 @@ static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y,
 static void ggml_mul_mat_vec_nc_f16_f32_sycl(
     const void *vx, const float *y, float *dst, const int ncols_x,
     const int nrows_x, const int row_stride_x, const int nchannels_x,
-    const int nchannels_y, const int channel_stride_x, queue_ptr stream) {
+    const int nchannels_y, const int channel_stride_x, const int channel_stride_y, queue_ptr stream) {
 
     const sycl::range<3> block_nums(nchannels_y, nrows_x, 1);
     const sycl::range<3> block_dims(1, 1, WARP_SIZE);
@@ -1834,7 +1834,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x,
-                                       row_stride_x, channel_stride_x,
+                                       row_stride_x, channel_stride_x, channel_stride_y,
                                        nchannels_y / nchannels_x, item_ct1);
             });
     }
@@ -1842,7 +1842,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl(
 
 
 
-static void scale_f32_sycl(const float *x, float *dst, const float scale,
+static void scale_f32_sycl(const float *x, float *dst, const float scale, const float bias,
                            const int k, queue_ptr stream) {
     const int num_blocks = (k + SYCL_SCALE_BLOCK_SIZE - 1) / SYCL_SCALE_BLOCK_SIZE;
     stream->parallel_for(
@@ -1850,7 +1850,7 @@ static void scale_f32_sycl(const float *x, float *dst, const float scale,
                               sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE)),
         [=](sycl::nd_item<3> item_ct1) {
-            scale_f32(x, dst, scale, k, item_ct1);
+            scale_f32(x, dst, scale, bias, k, item_ct1);
         });
 }
 
@@ -2123,8 +2123,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
 #if GGML_SYCL_DNNL
         if (!g_ggml_sycl_disable_dnn) {
-            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ptr,
-                                      DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
+                DnnlGemmWrapper::row_gemm(ctx,row_diff, src1_ncols , ne10, src0_ptr,
+                                     DnnlGemmWrapper::to_dt<sycl::half>(), src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
                                       dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
@@ -2170,8 +2170,8 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
 #if GGML_SYCL_DNNL
         if (!g_ggml_sycl_disable_dnn) {
-            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ddf1_i,
-                                      DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
+            DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,
+                                      DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
                                       dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
@@ -2319,9 +2319,11 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor * ds
     float *       dst_dd  = static_cast<float *>(dst->data);
 
     float scale;
-    memcpy(&scale, dst->op_params, sizeof(float));
+    float bias;
+    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&bias,  (float *) dst->op_params + 1, sizeof(float));
 
-    scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(dst->src[0]), main_stream);
+    scale_f32_sycl(src0_dd, dst_dd, scale, bias, ggml_nelements(dst->src[0]), main_stream);
     /*
     DPCT1010:87: SYCL uses exceptions to report errors and does not use the
     error codes. The call was replaced with 0. You need to rewrite this code.
@@ -2773,6 +2775,7 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
     const int64_t nb02 = src0->nb[2];
 
     const int64_t ne12 = src1->ne[2];
+    const int64_t nb11 = src1->nb[1];
 
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     queue_ptr main_stream = ctx.stream();
@@ -2783,8 +2786,9 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
 
     const int64_t row_stride_x = nb01 / sizeof(sycl::half);
     const int64_t channel_stride_x = nb02 / sizeof(sycl::half);
+    const int64_t channel_stride_y = nb11 / sizeof(float);
 
-    ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
+    ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x,channel_stride_y, main_stream);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -2838,8 +2842,8 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     float *            dst_ddf  = static_cast<float *>(dst->data);
 
     const sycl::half * src1_f16       = static_cast<const sycl::half *>(src1->data);
+    const size_t       type_size_src0 = ggml_type_size(src0->type);
     const size_t       type_size_src1 = ggml_type_size(src1->type);
-    GGML_ASSERT(nb10 == type_size_src1);
 
     // SRC1 strides
     int64_t                          s11 = nb11 / type_size_src1;
@@ -2851,11 +2855,40 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     if (src1->type != GGML_TYPE_F16) {
         scope_op_debug_print    scope_dbg_print(__func__, "/to_fp16_nc_sycl", dst, /*num_src=*/2,
                                                 " : converting src1 to fp16");
-        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
-        GGML_ASSERT(to_fp16_nc_sycl != nullptr);
+
+        // iterate tensor dims and find the slowest moving dim and stride
+        int64_t last_dim=0;
+        int64_t last_str=0;
+        int64_t largest_str=0;
+        for(int i = 0; i< 4; i++){
+            // last stride is always the largest
+            if(src1->nb[i] == largest_str){
+                if(src1->ne[last_dim] == 1){
+                    last_str = i;
+                    last_dim = i;
+                }
+            }
+            if(src1->nb[i] > largest_str){
+                largest_str = src1->nb[i];
+                last_str = i;
+                last_dim = i;
+            }
+
+        }
+#if GGML_SYCL_DNNL
+        // oneDNN handles strided data and does not need overhead of get_to_fp16_nc_sycl
+        const int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;
+        src1_f16_alloc.alloc(ne_src1);
+        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
+        GGML_ASSERT(to_fp16_sycl != nullptr);
+        to_fp16_sycl(src1_f16, src1_f16_alloc.get(), ne_src1, queue);
+# else
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
+        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
+        GGML_ASSERT(to_fp16_nc_sycl != nullptr);
         to_fp16_nc_sycl(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, queue);
+#endif
 
         src1_f16 = src1_f16_alloc.get();
         s11      = ne10;
@@ -2889,38 +2922,89 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
 
 #if GGML_SYCL_DNNL
     if (!g_ggml_sycl_disable_dnn) {
-        auto dnn_gemm = [&ctx, queue, ne11, ne01, ne10, nb00, nb01, nb02, s11, s12]
-            (const sycl::half* src1, const sycl::half* src0, float* dst, const dnnl_dim_t batches_a, const dnnl_dim_t batches_b) {
+            int64_t str_a0 = nb00 / type_size_src0;
+            int64_t str_a1 = nb01 / type_size_src0;
+            int64_t str_a2 = nb02 / type_size_src0;
 
-            DnnlGemmWrapper::gemm(ctx, ne11,ne01, ne10,
-                            src1, DnnlGemmWrapper::to_dt<sycl::half>(), s11, 1, s12,
-                            src0, DnnlGemmWrapper::to_dt<sycl::half>(), 1, nb01/nb00, nb02/nb00,
-                            dst, DnnlGemmWrapper::to_dt<float>(), queue, batches_a, batches_b);
-        };
+            int64_t str_b0 = nb10 / type_size_src1;
+            int64_t str_b1 = nb11 / type_size_src1;
+            int64_t str_b2 = nb12 / type_size_src1;
 
-        if (r2 == 1 && r3 == 1) {
-            if (ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
-                dnn_gemm(src1_f16, src0_f16, dst_ddf, ne12*ne13, ne02 * ne03);
-            }
-            else {
-                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
-                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie03*nb03)/sizeof(sycl::half)); // nb is in bytes
-                    const sycl::half* src1_f16_shifted = src1_f16 + ie03*s13;
-                    float* dst_shifted = dst_ddf + ((ie03*nb3)/sizeof(float));
-                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, ne12, ne02);
+            auto launch_gemm_for_batches = [&ctx, queue](const sycl::half *src0,
+                                                const sycl::half *src1, float *dst,
+                                                int64_t a0, int64_t a1, int64_t batcha,
+                                                int64_t b0, int64_t b1, int64_t batchb,
+                                                int64_t sa0, int64_t sa1, int64_t sa2,
+                                                int64_t sb0, int64_t sb1, int64_t sb2,
+                                                int64_t sd2) {
+                bool supported_broadcast = batchb == batcha ? true
+                        : batchb == 1 || batcha == 1        ? true
+                                                            : false;
+                if (supported_broadcast) {
+                    DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0,
+                            DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2, src1,
+                            DnnlGemmWrapper::to_dt<sycl::half>(), sb0, sb1, sb2, dst,
+                            DnnlGemmWrapper::to_dt<float>(), queue, batcha, batchb);
+                } else {
+                    // iterate over batches from smaller set of matrices (matrix 0)
+                    int64_t batches0 = batcha;
+                    int64_t batches1 = batchb;
+
+                    if (batches0 > batches1) {
+                        int64_t num_mul_mats = batches1;
+                        int64_t sub_batch = batches0 / num_mul_mats;
+                        // src0 is batched and bigger, shift and multiply with src1
+                        for (int64_t i0 = 0; i0 < num_mul_mats; i0++) {
+                            const sycl::half *src0_shifted = src0 + (sa2 * i0 * sub_batch);
+                            const sycl::half *src1_shifted = src1 + (sb2 * i0);
+                            float *dst_shifted = dst + (sd2 * i0 * sub_batch);
+                            DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0_shifted,
+                                    DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2,
+                                    src1_shifted, DnnlGemmWrapper::to_dt<sycl::half>(), sb0,
+                                    sb1, sb2, dst_shifted, DnnlGemmWrapper::to_dt<float>(),
+                                    queue, sub_batch, 1);
+                        }
+                    } else {
+                        int64_t num_mul_mats = batches0;
+                        int64_t sub_batch = batches1 / num_mul_mats;
+                        // src1 is batched and bigger, shift and multiply with src0
+                        for (int64_t i1 = 0; i1 < num_mul_mats; i1++) {
+                            const sycl::half *src0_shifted = src0 + (sa2 * i1);
+                            const sycl::half *src1_shifted = src1 + (sb2 * i1 * sub_batch);
+                            float *dst_shifted = dst + (sd2 * i1 * sub_batch);
+                            DnnlGemmWrapper::gemm(ctx, a1, b1, a0, src0_shifted,
+                                    DnnlGemmWrapper::to_dt<sycl::half>(), sa0, sa1, sa2,
+                                    src1_shifted, DnnlGemmWrapper::to_dt<sycl::half>(), sb0,
+                                    sb1, sb2, dst_shifted, DnnlGemmWrapper::to_dt<float>(),
+                                    queue, 1, sub_batch);
+                        }
+                    }
+                }
+            };
+
+            bool cont_batches_a = nb02 * ne02 == nb03;
+            bool cont_batches_b = nb12 * ne12 == nb13;
+            if (cont_batches_a && cont_batches_b) {
+                int64_t batches0 = ne02 * ne03;
+                int64_t batches1 = ne12 * ne13;
+                launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
+                        ne10, ne11, batches1, str_a0, str_a1, str_a2, str_b0, str_b1,
+                        str_b2, nb2 / sizeof(float));
+            } else {
+                for (int64_t b_a = 0; b_a < ne03; b_a++) {
+                    const sycl::half *src0_f16_shifted
+                            = src0_f16 + (nb03 * b_a / type_size_src0);
+                    const sycl::half *src1_f16_shifted
+                            = src1_f16 + (nb13 * b_a / type_size_src1);
+                    float *dst_shifted = dst_ddf + (nb3 * b_a / sizeof(float));
+                    int64_t batches0 = ne02;
+                    int64_t batches1 = ne12;
+                    launch_gemm_for_batches(src0_f16_shifted, src1_f16_shifted, dst_shifted,
+                            ne00, ne01, batches0, ne10, ne11, batches1, str_a0, str_a1,
+                            str_a2, str_b0, str_b1, str_b2, nb2 / sizeof(float));
                 }
             }
-        } else {
-            // iterate over batches from smaller set of matrices (matrix 0)
-            for (int64_t ie02 = 0; ie02 < ne02; ++ie02) {
-                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
-                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie02*nb02 + ie03*nb03)/sizeof(sycl::half));
-                    const sycl::half* src1_f16_shifted = src1_f16 + ie02*s12*r2 + ie03*s13*r3;
-                    float* dst_shifted = dst_ddf + ((ie02*nb2*r2 + ie03*nb3*r3)/sizeof(float));
-                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, r2*r3, 1);
-                }
-            }
-        }
+
     }
     else
 #endif
@@ -3260,10 +3344,10 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
             // The kernel from the if path is faster for that specific case, but does not support all mul mats.
             ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
         }
-    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2] * src1->ne[3] > 1) {
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
@@ -3446,8 +3530,11 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             SYCL_CHECK(CHECK_TRY_ERROR(
                 stream->memset(dev_cur_src1_row.get(), 0, sizeof(int))));
 
+            const unsigned int max_work_group_size = ggml_sycl_info().max_work_group_sizes[ctx.device];
+            assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0);
+
             {
-                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u));
+                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, max_work_group_size));
                 sycl::range<3> grid_dims(1, n_ids, ids->ne[1]);
                 sycl_launch(stream, [&](sycl::handler & cgh) {
                     sycl::local_accessor<int, 0> src1_row_acc(cgh);
@@ -3491,7 +3578,7 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             ggml_sycl_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
 
             {
-                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u));
+                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, max_work_group_size));
                 sycl::range<3> grid_dims(1, 1, num_src1_rows);
                 sycl_launch(stream, [&](sycl::handler & cgh) {
                     const char *__restrict dst_contiguous_get =
@@ -3603,6 +3690,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_GET_ROWS:
             ggml_sycl_get_rows(ctx, dst);
             break;
+        case GGML_OP_SET_ROWS:
+            ggml_sycl_op_set_rows(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_sycl_dup(ctx, dst);
             break;
@@ -4297,7 +4387,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 // TODO: add support
                 // ref: https://github.com/ggml-org/llama.cpp/pull/14274
-                return false;
+#pragma message("TODO: implement BF16, 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->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_I64));
             } break;
         case GGML_OP_CPY:
             {

ggml/src/ggml-sycl/im2col.cpp

@@ -26,7 +26,7 @@ static void im2col_kernel(const float * x, T * dst, int64_t batch_offset, int64_
 
     // make each work-item deal with more elements since sycl global range can not exceed max int
     for (int64_t i = global_id; i < pelements; i += (work_group_size * item_ct1.get_group_range(2))) {
-        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-sycl/set_rows.cpp

@@ -0,0 +1,131 @@
+#include "set_rows.hpp"
+
+namespace utils {
+template<typename T>
+static constexpr bool is_arithmetic_v() {
+    return std::is_arithmetic_v<T> || std::is_same_v<T, sycl::half> || std::is_same_v<T, sycl::ext::oneapi::bfloat16>;
+}
+}
+
+template<typename TIn, typename TOut>
+static inline std::enable_if_t<utils::is_arithmetic_v<TIn>() && utils::is_arithmetic_v<TOut>(), void>
+convert (const char* src, char* dst) {
+    auto src_val = *reinterpret_cast<const TIn*>(src);
+    auto dst_val = sycl::vec<TIn, 1>(src_val).template convert<TOut, sycl::rounding_mode::automatic>()[0];
+   *reinterpret_cast<TOut*>(dst) = dst_val;
+}
+
+template<typename TIn, typename TOut>
+static void k_set_rows(
+        const char * __restrict__ src0, const int64_t * __restrict__ src1, char * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t ne11, const int64_t ne12,
+        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,
+        const size_t src_type_size, const size_t dst_type_size,
+        const int64_t total_elements,
+        const sycl::nd_item<1> & item_ct1) {
+
+    const int64_t i = item_ct1.get_global_linear_id();
+    if (i >= total_elements) {
+        return;
+    }
+
+    const int64_t i03 = i / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i - 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 = *(const int64_t *)((const char *)src1 + calculate_offset<3>({nb10, nb11, nb12}, {i10, i11, i12}));
+
+    const char * src0_row = src0 + calculate_offset<3>({nb01, nb02, nb03}, {i01, i02, i03});
+    const char * src_elem = src0_row + i00 * src_type_size;
+    char * dst_row_ptr = dst + dst_row*nb1 + i02*nb2 + i03*nb3;
+    char * dst_elem = dst_row_ptr + i00 * dst_type_size;
+
+    convert<TIn, TOut>(src_elem, dst_elem);
+}
+
+template<typename TIn, typename TOut>
+static void set_rows_sycl(
+        const char * src0_d, const int64_t * src1_d, char * dst_d,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne11, const int64_t ne12, 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,
+        const size_t src_type_size, const size_t dst_type_size,
+        queue_ptr stream) {
+
+    const int64_t total_elements = ne00 * ne01 * ne02 * ne03;
+
+    constexpr int block_size = 64;
+    const int64_t grid_size = ceil_div(total_elements, block_size);
+
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<1>(grid_size * block_size, block_size),
+        [=](sycl::nd_item<1> item_ct1) {
+            k_set_rows<TIn, TOut>(
+                src0_d, src1_d, dst_d,
+                ne00, ne01, ne02,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                src_type_size, dst_type_size,
+                total_elements,
+                item_ct1
+            );
+        }
+    );
+}
+
+void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I64);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t * src1_dd = static_cast<const int64_t *>(src1->data);
+
+    dpct::queue_ptr stream = ctx.stream();
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            set_rows_sycl<float, float>(
+                (const char *)src0->data, src1_dd, (char *)dst->data,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                sizeof(float), sizeof(float),
+                stream
+            );
+            break;
+        case GGML_TYPE_F16:
+            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+            set_rows_sycl<float, sycl::half>(
+                (const char *)src0->data, src1_dd, (char *)dst->data,
+                ne00, ne01, ne02, ne03,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                sizeof(float), sizeof(sycl::half),
+                stream
+            );
+            break;
+        default:
+            GGML_ABORT("Unsupported tensor type!");
+            break;
+    }
+}

ggml/src/ggml-sycl/set_rows.hpp

@@ -0,0 +1,8 @@
+#ifndef GGML_SYCL_SET_ROWS_HPP
+#define GGML_SYCL_SET_ROWS_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_SET_ROWS_HPP

ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -0,0 +1,265 @@
+#version 450
+
+#ifdef USE_COLLECTIVES
+#    extension GL_KHR_shader_subgroup_shuffle : enable
+#endif
+
+#include "types.comp"
+
+// Make spec constant
+#define SHMEM_PAD 0
+
+// shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
+layout(binding = 0) readonly buffer A {
+    A_TYPE knl_data[];
+};  // src0 - kernel:   [KW, KH, Cin, Cout]
+
+layout(binding = 1) readonly buffer B {
+    B_TYPE src_data[];
+};  // src1 - input:    [W, H, Cin, N] -- channel_first format
+
+layout(binding = 2) writeonly buffer D {
+    D_TYPE dst_data[];
+};  // dst - result:    [OW, OH, Cout, N]
+
+layout(push_constant) uniform parameter {
+    // I/O channels, batch size
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t N;
+
+    // Tensor spatial sizes: kernel, input, output
+    uint32_t KW;
+    uint32_t KH;
+    uint32_t W;
+    uint32_t H;
+    uint32_t OW;
+    uint32_t OH;
+
+    // Parameters: stride, padding, dilation - 0=y, 1=x
+    uint32_t s0;
+    uint32_t s1;
+    uint32_t p0;
+    uint32_t p1;
+    uint32_t d0;
+    uint32_t d1;
+
+    // Strides in elements
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+}
+
+p;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+// Blocktile sizes
+layout(constant_id = 1) const uint BS_K            = 128;
+layout(constant_id = 2) const uint BS_CRS          = 16;
+layout(constant_id = 3) const uint BS_NPQ          = 128;
+// Thread-tile sizes
+layout(constant_id = 4) const uint TS_K            = 8;
+layout(constant_id = 5) const uint use_collectives = 1;
+
+uint32_t       tid     = gl_LocalInvocationID.x;
+const uint32_t WG_SIZE = gl_WorkGroupSize.x;
+
+uint splitWork(uint work_size, uint block_size) {
+    return (block_size + work_size - 1) / block_size;
+}
+
+uint32_t K   = p.Cout;
+uint32_t CRS = p.Cin * p.KH * p.KW;
+uint32_t NPQ = p.N * p.OH * p.OW;
+
+uint32_t n_elems_out = K * NPQ;
+
+// Number of blocktiles per input
+uint32_t NB_CRS = splitWork(CRS, BS_CRS);
+
+const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
+const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
+
+const uint32_t Ash_numel = BS_K * BS_CRS;
+const uint32_t Bsh_numel = BS_CRS * BS_NPQ;
+
+const uint32_t Ash_len = BS_K * Ash_stride;
+const uint32_t Bsh_len = BS_CRS * Bsh_stride;
+
+shared float Ash[Ash_len];  // K x CRS
+shared float Bsh[Bsh_len];  // CRS x NPQ
+
+// Threadtile sizes
+const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
+
+// Number of threadtiles per blocktile
+const uint32_t NT_K   = BS_K / TS_K;
+const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
+
+float regA[TS_K];
+float regB[TS_NPQ];
+float regC[TS_K][TS_NPQ];
+
+/*
+Compute
+KxCRS @ CRSxNPQ = K x NPQ
+K=Cout
+C=Cin
+R,S=KH,KW
+P,Q=OH,OW
+*/
+
+uint32_t B_idx_K   = gl_WorkGroupID.x;
+uint32_t B_idx_NPQ = gl_WorkGroupID.y;
+
+uint32_t T_y = tid / NT_NPQ;
+uint32_t T_x = tid % NT_NPQ;
+
+uint32_t       Ar    = tid / BS_CRS;
+uint32_t       Ac    = tid % BS_CRS;
+const uint32_t ArpWg = WG_SIZE / BS_CRS;
+
+uint32_t       Br    = tid / BS_NPQ;
+uint32_t       Bc    = tid % BS_NPQ;
+const uint32_t BrpWg = WG_SIZE / BS_NPQ;
+
+void main() {
+    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+            regC[T_ly][T_lx] = 0.0;
+        }
+    }
+    /* Advance block in CRS dim */
+    for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
+        uint32_t CRS_idx_a;
+        uint32_t Cin_idx_a;
+        uint32_t KH_idx_a;
+        uint32_t KW_idx_a;
+
+#ifdef USE_COLLECTIVES
+        uint32_t cached_CRS_idx;
+        uint32_t cached_Cin_idx;
+        uint32_t cached_KH_idx;
+        uint32_t cached_KW_idx;
+        if (use_collectives == 1) {
+            cached_CRS_idx                = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
+            cached_Cin_idx                = cached_CRS_idx / (p.KW * p.KH);
+            uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
+            cached_KH_idx                 = cached_CRS_remainder / p.KW;
+            cached_KW_idx                 = cached_CRS_remainder - cached_KH_idx * p.KW;
+
+            CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
+            Cin_idx_a = subgroupShuffle(cached_Cin_idx, Ac);
+            KH_idx_a  = subgroupShuffle(cached_KH_idx, Ac);
+            KW_idx_a  = subgroupShuffle(cached_KW_idx, Ac);
+        } else {
+            CRS_idx_a              = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
+            Cin_idx_a              = CRS_idx_a / (p.KW * p.KH);
+            uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
+            KH_idx_a               = CRS_remainder / p.KW;
+            KW_idx_a               = CRS_remainder - KH_idx_a * p.KW;
+        }
+#else
+        CRS_idx_a     = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
+        Cin_idx_a     = CRS_idx_a / (p.KW * p.KH);
+        CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
+        KH_idx_a      = CRS_remainder / p.KW;
+        KW_idx_a      = CRS_remainder - KH_idx_a * p.KW;
+#endif
+
+        /* Load kernel to A_block: (BS_K x BS_CRS)*/
+        for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
+            uint32_t B_ly    = r_offset + Ar;
+            uint32_t B_lx    = Ac;
+            uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
+            uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + Cin_idx_a * p.nb02 + K_idx * p.nb03, K * CRS - 1);
+            float    val     = knl_data[knl_idx];
+            if (K_idx >= K || CRS_idx_a >= CRS) {
+                val = 0.0;
+            }
+            Ash[B_ly * Ash_stride + B_lx] = val;
+        }
+        /* Load input to B_block: (BS_CRS x BS_NPQ) */
+        for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+            uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
+            uint32_t B_lx          = Bc;
+            uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
+            uint32_t N_idx         = NPQ_idx / (p.OH * p.OW);
+            uint32_t NPQ_remainder = NPQ_idx - N_idx * p.OH * p.OW;
+            uint32_t OH_idx        = NPQ_remainder / p.OW;
+            uint32_t OW_idx        = NPQ_remainder - OH_idx * p.OW;
+
+            uint32_t CRS_idx_b;
+            uint32_t Cin_idx_b;
+            uint32_t KH_idx_b;
+            uint32_t KW_idx_b;
+#ifdef USE_COLLECTIVES
+            if (use_collectives == 1) {
+                CRS_idx_b = subgroupShuffle(cached_CRS_idx, r_offset + Br);
+                Cin_idx_b = subgroupShuffle(cached_Cin_idx, r_offset + Br);
+                KH_idx_b  = subgroupShuffle(cached_KH_idx, r_offset + Br);
+                KW_idx_b  = subgroupShuffle(cached_KW_idx, r_offset + Br);
+            } else {
+                CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
+                Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+                uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
+                KH_idx_b               = CRS_remainder / p.KW;
+                KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+            }
+#else
+            CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
+            Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+            uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
+            KH_idx_b               = CRS_remainder / p.KW;
+            KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+#endif
+
+            uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
+            uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
+            uint32_t src_idx =
+                min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
+            float val = src_data[src_idx];
+            if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
+                val = 0.0;
+            }
+            Bsh[B_ly * Bsh_stride + B_lx] = val;
+        }
+        barrier();
+        for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
+            }
+            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
+            }
+            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                    regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                }
+            }
+        }
+        barrier();
+    }
+    /* Save C* */
+    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+            uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
+            uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
+            uint32_t N_idx   = NPQ_idx / (p.OH * p.OW);
+            uint32_t OH_idx  = (NPQ_idx - N_idx * p.OH * p.OW) / p.OW;
+            uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+            uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+            if (K_idx < K && NPQ_idx < NPQ) {
+                dst_data[dst_idx] = regC[T_ly][T_lx];
+            }
+        }
+    }
+}

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

@@ -1,22 +1,26 @@
 #version 450
 
-#if RTE16
-#extension GL_EXT_spirv_intrinsics : enable
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif // RTE16
-
+#include "rte.comp"
 #include "types.comp"
-#include "generic_unary_head.comp"
 
-#if defined(DATA_A_IQ4_NL)
-// 16 invocations needed for init_iq4nl_shmem
-layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
+#if defined(SET_ROWS) && QUANT_K == 1
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 512;
 #else
-layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 32;
 #endif
 
 layout (binding = 0) readonly buffer S {float data_s[];};
+
+#if defined(SET_ROWS)
+#include "generic_binary_head.comp"
+layout (binding = 1) readonly buffer C {uvec2 data_i[];};
+layout (binding = 2) writeonly buffer Q {A_TYPE data_q[];};
+#else
+#include "generic_unary_head.comp"
 layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
+#endif
 
 #if defined(DATA_A_Q4_0)
 void quantize(uint dst_idx, uint src_idx)
@@ -221,15 +225,56 @@ void quantize(uint dst_idx, uint src_idx)
 }
 #endif
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(data_s[src_idx]);
+}
+#endif
+
+#if defined(DATA_A_BF16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(fp32_to_bf16(data_s[src_idx]));
+}
+#endif
+
+#if defined(SET_ROWS)
+
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
-    if (gl_LocalInvocationIndex.x != 0) {
-        return;
-    }
 #endif
 
-    const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
+    const uint idx = ((gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x) * BLOCK_SIZE + gl_LocalInvocationID.x) * QUANT_K;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    uint i00, i01, i02, i03;
+    get_indices(idx, i00, i01, i02, i03);
+
+    uint i12 = fastmod(i03, p.ne12);
+    uint i11 = fastmod(i02, p.ne11);
+    uint i10 = i01;
+
+    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()].x;
+
+    uint src0_idx = src0_idx(i00, i01, i02, i03) + get_aoffset();
+    uint dst_idx = dst_idx(i00 / QUANT_K, i1, i02, i03) + get_doffset();
+
+    quantize(dst_idx, src0_idx);
+}
+
+#else
+
+void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
+    const uint idx = (gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x) * QUANT_K;
 
     if (idx >= p.ne) {
         return;
@@ -240,3 +285,5 @@ void main() {
 
     quantize(dst_idx, src_idx);
 }
+
+#endif

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = uint(gl_WorkGroupID.x * 256 + wgy);
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
         const uint tid = gl_LocalInvocationID.x;

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

@@ -1,6 +1,8 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_control_flow_attributes : require
 
+#include "rte.comp"
+
 layout (push_constant) uniform parameter
 {
     uint ne;

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

@@ -1,5 +1,7 @@
 #extension GL_EXT_shader_16bit_storage : require
 
+#include "rte.comp"
+
 layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};

ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -1,12 +1,9 @@
 #version 450
 
 #extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_spirv_intrinsics: enable
 #extension GL_EXT_control_flow_attributes : require
 
-#if RTE16
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif
+#include "rte.comp"
 
 layout (push_constant) uniform parameter
 {
@@ -43,12 +40,10 @@ void main() {
     const uint src_base = ic * p.offset_delta + batch * p.batch_offset;
     const uint dst_base = ((batch * p.OH + oh) * p.OW) * p.CHW + ic * (p.KW * p.KH);
     const int oh_s1 = int(oh) * p.s1;
-    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint ksize = p.OW * p.KH;
 
     const uint base_linear_idx = gidx * NUM_ITER;
 
-    const uint max_ky = ksize / p.OW;
-
     uint current_kx = base_linear_idx / ksize;
     const uint rem = base_linear_idx - (current_kx * ksize);
     uint current_ky = rem / p.OW;
@@ -79,7 +74,7 @@ void main() {
 
         if (++current_ix == p.OW) {
             current_ix = 0;
-            if (++current_ky == max_ky) {
+            if (++current_ky == p.KH) {
                 current_ky = 0;
                 current_kx++;
             }

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

@@ -18,6 +18,7 @@
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
 #endif
 
 #ifdef MUL_MAT_ID
@@ -104,6 +105,10 @@ shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
 
 #ifdef MUL_MAT_ID
 shared u16vec2 row_ids[4096];
+uint _ne1;
+#ifdef COOPMAT
+shared uint _ne1_sh;
+#endif
 #endif // MUL_MAT_ID
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)
@@ -172,7 +177,47 @@ void main() {
     const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
 
 #ifdef MUL_MAT_ID
-    uint _ne1 = 0;
+#ifdef COOPMAT
+    // Spread the search across all elements in the first subgroup
+    if (gl_SubgroupID == 0) {
+        _ne1 = 0;
+        uint num_elements = p.nei1 * p.nei0;
+
+        uint ids[16];
+        uint iter = 0;
+
+        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
+            // prefetch up to 16 elements
+            if (iter == 0) {
+                [[unroll]] for (uint k = 0; k < 16; ++k) {
+                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
+                    bool in_range = i < num_elements;
+                    uint ii1 = i / p.nei0;
+                    uint ii0 = i % p.nei0;
+                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+                }
+            }
+            uint i = j + gl_SubgroupInvocationID;
+            bool in_range = i < num_elements;
+            uint ii1 = i / p.nei0;
+            uint ii0 = i % p.nei0;
+            uint id = ids[iter++];
+            uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+            uint idx = subgroupBallotExclusiveBitCount(ballot);
+            if (in_range && id == expert_idx) {
+                row_ids[_ne1 + idx] = u16vec2(ii0, ii1);
+            }
+            _ne1 += subgroupBallotBitCount(ballot);
+            iter &= 15;
+        }
+        _ne1_sh = _ne1;
+    }
+
+    barrier();
+
+    _ne1 = _ne1_sh;
+#else
+    _ne1 = 0;
     for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
         for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
             if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
@@ -183,6 +228,7 @@ void main() {
     }
 
     barrier();
+#endif
 
     // Workgroup has no work
     if (ic * BN >= _ne1) return;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -162,17 +162,32 @@ void main() {
         _ne1 = 0;
         uint num_elements = p.nei1 * p.nei0;
 
-        for (uint i = gl_SubgroupInvocationID; subgroupAny(i < num_elements); i += gl_SubgroupSize) {
+        uint ids[16];
+        uint iter = 0;
+
+        for (uint j = 0; j < num_elements; j += gl_SubgroupSize) {
+            // prefetch up to 16 elements
+            if (iter == 0) {
+                [[unroll]] for (uint k = 0; k < 16; ++k) {
+                    uint i = j + gl_SubgroupInvocationID + k*gl_SubgroupSize;
+                    bool in_range = i < num_elements;
+                    uint ii1 = i / p.nei0;
+                    uint ii0 = i % p.nei0;
+                    ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+                }
+            }
+            uint i = j + gl_SubgroupInvocationID;
             bool in_range = i < num_elements;
-            uint ii0 = i % p.nei0;
             uint ii1 = i / p.nei0;
-            uint id = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            uint ii0 = i % p.nei0;
+            uint id = ids[iter++];
             uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
             uint idx = subgroupBallotExclusiveBitCount(ballot);
             if (in_range && id == expert_idx) {
                 row_ids[_ne1 + idx] = u16vec4(ii0 % p.ne11, ii1, ii0, 0);
             }
             _ne1 += subgroupBallotBitCount(ballot);
+            iter &= 15;
         }
         _ne1_sh = _ne1;
     }
@@ -414,17 +429,31 @@ void main() {
                 fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
             }
 
-            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+            if ((ir + 1) * BM <= p.M && block_k + BK <= end_k) {
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
-            coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID
-            coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
 #else
-            coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
 #endif
 
-            sum = coopMatMulAdd(mat_a, mat_b, sum);
+                sum = coopMatMulAdd(mat_a, mat_b, sum);
+            } else {
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
+#ifdef MUL_MAT_ID
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+#else
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
+#endif
+
+                sum = coopMatMulAdd(mat_a, mat_b, sum);
+            }
         }
 
         // Convert from ACC_TYPE to D_TYPE

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

@@ -0,0 +1,46 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+uint wrap_idx(int i, uint ne) {
+    if (i < 0) {
+        return i + ne;
+    } else if (i >= ne) {
+        return i - ne;
+    }
+    return i;
+}
+
+void main() {
+    const uint idx = get_idx();
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
+    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i2_offset = i2*p.ne11*p.ne10;
+    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
+
+    const uint p1 = floatBitsToUint(p.param1);
+    const uint p2 = floatBitsToUint(p.param2);
+    const int s0 = int(p1 >> 16)    - 0x8000;
+    const int s1 = int(p1 & 0xFFFF) - 0x8000;
+    const int s2 = int(p2 >> 16)    - 0x8000;
+    const int s3 = int(p2 & 0xFFFF) - 0x8000;
+
+    const uint i00 = wrap_idx(int(i0) - s0, p.ne10);
+    const uint i01 = wrap_idx(int(i1) - s1, p.ne11);
+    const uint i02 = wrap_idx(int(i2) - s2, p.ne12);
+    const uint i03 = wrap_idx(int(i3) - s3, p.ne13);
+
+    const uint a_idx = i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
+    const uint d_idx = i3 *p.nb13 + i2 *p.nb12 + i1 *p.nb11 + i0 *p.nb10;
+
+    data_d[get_doffset() + d_idx] = D_TYPE(data_a[get_aoffset() + a_idx]);
+}

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

@@ -1,11 +1,8 @@
 #include "types.comp"
 
 #extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_spirv_intrinsics: enable
 
-#if RTE16
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif
+#include "rte.comp"
 
 layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
 

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

@@ -0,0 +1,5 @@
+
+#if RTE16
+#extension GL_EXT_spirv_intrinsics : enable
+spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
+#endif // RTE16

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

@@ -18,7 +18,7 @@ void main() {
             continue;
         }
 
-        data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) * FLOAT_TYPE(p.param1));
+        data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) * FLOAT_TYPE(p.param1) + FLOAT_TYPE(p.param2));
         idx += num_threads;
     }
 }

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

@@ -3,6 +3,7 @@
 layout (push_constant) uniform parameter
 {
     uint ne; uint a_offset; uint d_offset;
+    uint ne00; uint ne01;
     uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -15,6 +16,61 @@ layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
+// from ggml.h: enum ggml_scale_mode, enum ggml_scale_flag
+#define NEAREST  0
+#define BILINEAR 1
+#define ALIGN_CORNERS (1 << 8)
+
+layout (constant_id = 0) const uint scale_mode = 0;
+
+float fetch_nearest(uint i10, uint i11, uint i12, uint i13) {
+    const uint i00 = uint(i10 / p.sf0);
+    const uint i01 = uint(i11 / p.sf1);
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+
+    return data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00];
+}
+
+float fetch_bilinear(ivec2 c0, ivec2 c1, vec2 d, uint i12, uint i13) {
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+    const uint base = p.a_offset + i03 * p.nb03 + i02 * p.nb02;
+
+    const float v00 = data_a[base + c0.y * p.nb01 + c0.x * p.nb00];
+    const float v01 = data_a[base + c0.y * p.nb01 + c1.x * p.nb00];
+    const float v10 = data_a[base + c1.y * p.nb01 + c0.x * p.nb00];
+    const float v11 = data_a[base + c1.y * p.nb01 + c1.x * p.nb00];
+
+    return
+        v00 * (1.0-d.x) * (1.0-d.y) +
+        v01 * d.x       * (1.0-d.y) +
+        v10 * (1.0-d.x) * d.y +
+        v11 * d.x       * d.y;
+}
+
+float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
+    const ivec2 ne0 = ivec2(p.ne00, p.ne01);
+
+    const vec2 c = (vec2(i10, i11) + 0.5) / vec2(p.sf0, p.sf1) - 0.5;
+    const vec2 c0f = floor(c);
+    const vec2 d = c - c0f;
+    const ivec2 c0 = max(ivec2(c0f), 0);
+    const ivec2 c1 = min(ivec2(c0f + 1), ne0 - 1);
+
+    return fetch_bilinear(c0, c1, d, i12, i13);
+}
+
+float interpolate_bilinear_align_corners(uint i10, uint i11, uint i12, uint i13) {
+    const vec2 c = vec2(i10, i11) / vec2(p.sf0, p.sf1);
+    const vec2 c0f = floor(c);
+    const vec2 d = c - c0f;
+    const ivec2 c0 = ivec2(c0f);
+    const ivec2 c1 = c0 + 1;
+
+    return fetch_bilinear(c0, c1, d, i12, i13);
+}
+
 void main() {
     const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 
@@ -27,10 +83,18 @@ void main() {
     const uint i12 = (idx / (p.ne10 * p.ne11)) % p.ne12;
     const uint i13 = (idx / (p.ne10 * p.ne11 * p.ne12)) % p.ne13;
 
-    const uint i00 = uint(i10 / p.sf0);
-    const uint i01 = uint(i11 / p.sf1);
-    const uint i02 = uint(i12 / p.sf2);
-    const uint i03 = uint(i13 / p.sf3);
+    float result;
+    switch (scale_mode) {
+        case NEAREST:
+            result = fetch_nearest(i10, i11, i12, i13);
+            break;
+        case BILINEAR:
+            result = interpolate_bilinear(i10, i11, i12, i13);
+            break;
+        case BILINEAR | ALIGN_CORNERS:
+            result = interpolate_bilinear_align_corners(i10, i11, i12, i13);
+            break;
+    }
 
-    data_d[p.d_offset + idx] = D_TYPE(data_a[p.a_offset + i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+    data_d[p.d_offset + idx] = D_TYPE(result);
 }

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

@@ -518,6 +518,11 @@ void process_shaders() {
         string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }
 
+    for (std::string t : {"f32", "f16", "bf16", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
+        string_to_spv("set_rows_" + t, "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+        string_to_spv("set_rows_" + t + "_rte", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+    }
+
     auto get_type_str = [](bool f16) {
         return f16 ? "float16_t" : "float";
     };
@@ -532,8 +537,10 @@ void process_shaders() {
     for (auto src0_f16 : {false, true}) {
     for (auto src1_f16 : {false, true}) {
     for (auto dst_f16  : {false, true}) {
-        auto name = op + get_suffix(src0_f16, src1_f16, dst_f16);
-        string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}});
+    for (auto rte      : {false, true}) {
+        auto name = op + get_suffix(src0_f16, src1_f16, dst_f16) + (rte ? "_rte" : "");
+        string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
+    }
     }
     }
     }
@@ -587,16 +594,19 @@ void process_shaders() {
     string_to_spv("sigmoid_f16",    "sigmoid.comp",     {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("sigmoid_f32",    "sigmoid.comp",     {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
-    string_to_spv("geglu_f16",      "geglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("geglu_f32",      "geglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("reglu_f16",      "reglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("reglu_f32",      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("swiglu_f16",     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("swiglu_f32",     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("geglu_erf_f16",  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("geglu_erf_f32",  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("geglu_quick_f32","geglu_quick.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    for (auto rte : {false, true}) {
+        std::string suffix = rte ? "_rte" : "";
+        string_to_spv("geglu_f16" + suffix,      "geglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_f32" + suffix,      "geglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("reglu_f16" + suffix,      "reglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("reglu_f32" + suffix,      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_f16" + suffix,     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_f32" + suffix,     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_erf_f16" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_erf_f32" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_quick_f16" + suffix,"geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_quick_f32" + suffix,"geglu_quick.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+    }
 
     string_to_spv("leaky_relu_f32", "leaky_relu.comp",  {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("silu_back_f32",  "silu_back.comp",   {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -645,9 +655,13 @@ void process_shaders() {
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
+
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
 
+    string_to_spv("roll_f32", "roll.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+
     for (auto &c : compiles) {
         c.wait();
     }
@@ -702,11 +716,59 @@ void write_output_files() {
             std::remove(path.c_str());
         }
     }
+
+    std::string suffixes[2] = {"_f32", "_f16"};
     for (const char *op : {"add", "sub", "mul", "div"}) {
-        fprintf(hdr, "extern unsigned char *%s_data[2][2][2];\n", op);
-        fprintf(hdr, "extern uint64_t %s_len[2][2][2];\n", op);
-        fprintf(src, "unsigned char *%s_data[2][2][2] = {{{%s_f32_f32_f32_data, %s_f32_f32_f16_data}, {%s_f32_f16_f32_data, %s_f32_f16_f16_data}}, {{%s_f16_f32_f32_data, %s_f16_f32_f16_data}, {%s_f16_f16_f32_data, %s_f16_f16_f16_data}}};\n", op, op, op, op, op, op, op, op, op);
-        fprintf(src, "uint64_t %s_len[2][2][2] = {{{%s_f32_f32_f32_len, %s_f32_f32_f16_len}, {%s_f32_f16_f32_len, %s_f32_f16_f16_len}}, {{%s_f16_f32_f32_len, %s_f16_f32_f16_len}, {%s_f16_f16_f32_len, %s_f16_f16_f16_len}}};\n", op, op, op, op, op, op, op, op, op);
+        fprintf(hdr, "extern unsigned char *%s_data[2][2][2][2];\n", op);
+        fprintf(hdr, "extern uint64_t %s_len[2][2][2][2];\n", op);
+        std::string data = "unsigned char *" + std::string(op) + "_data[2][2][2][2] = ";
+        std::string len = "uint64_t " + std::string(op) + "_len[2][2][2][2] = ";
+        for (uint32_t t0 = 0; t0 < 2; ++t0) {
+            if (t0 == 0) {
+                data += "{";
+                len += "{";
+            }
+            for (uint32_t t1 = 0; t1 < 2; ++t1) {
+                if (t1 == 0) {
+                    data += "{";
+                    len += "{";
+                }
+                for (uint32_t t2 = 0; t2 < 2; ++t2) {
+                    if (t2 == 0) {
+                        data += "{";
+                        len += "{";
+                    }
+                    for (uint32_t rte = 0; rte < 2; ++rte) {
+                        if (rte == 0) {
+                            data += "{";
+                            len += "{";
+                        }
+                        data += op + suffixes[t0] + suffixes[t1] + suffixes[t2] + ((rte != 0) ? "_rte" : "");
+                        len  += op + suffixes[t0] + suffixes[t1] + suffixes[t2] + ((rte != 0) ? "_rte" : "");
+                        data += "_data,";
+                        len  += "_len,";
+                        if (rte == 1) {
+                            data += "}, ";
+                            len += "}, ";
+                        }
+                    }
+                    if (t2 == 1) {
+                        data += "}, ";
+                        len += "}, ";
+                    }
+                }
+                if (t1 == 1) {
+                    data += "}, ";
+                    len += "}, ";
+                }
+            }
+            if (t0 == 1) {
+                data += "};\n";
+                len += "};\n";
+            }
+        }
+        fputs(data.c_str(), src);
+        fputs(len.c_str(), src);
     }
     fclose(hdr);
     fclose(src);