CUDA: fix quantized KV cache + multiple sequences (#14822)

* CUDA: fix quantized KV cache + multiple sequences

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

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.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

CODEOWNERS

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

README.md

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

ci/run.sh

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

common/arg.cpp

@@ -1612,7 +1612,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.antiprompt.emplace_back(value);
         }
-    ).set_examples({LLAMA_EXAMPLE_MAIN}));
+    ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-sp", "--special"},
         string_format("special tokens output enabled (default: %s)", params.special ? "true" : "false"),
@@ -2655,6 +2655,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.i_chunk = value;
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
+    add_opt(common_arg(
+        {"--show-statistics"},
+        string_format("show imatrix statistics and then exit (default: %s)", params.show_statistics ? "true" : "false"),
+        [](common_params & params) {
+            params.show_statistics = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--parse-special"},
         string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"),

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();

common/common.h

@@ -432,9 +432,10 @@ struct common_params {
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
 
-    bool process_output = false; // collect data for the output tensor
-    bool compute_ppl    = true;  // whether to compute perplexity
-    bool parse_special  = false; // whether to parse special tokens during imatrix tokenization
+    bool process_output  = false; // collect data for the output tensor
+    bool compute_ppl     = true;  // whether to compute perplexity
+    bool show_statistics = false; // show imatrix statistics per tensor
+    bool parse_special   = false; // whether to parse special tokens during imatrix tokenization
 
     // cvector-generator params
     int n_pca_batch = 100;
@@ -534,6 +535,7 @@ static bool string_starts_with(const std::string & str,
 
 // While we wait for C++20's std::string::ends_with...
 bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
+bool string_remove_suffix(std::string & str, const std::string_view & suffix);
 size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);

convert_hf_to_gguf.py

@@ -843,6 +843,9 @@ class TextModel(ModelBase):
         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")
@@ -2861,7 +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.")
@@ -2894,6 +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",
@@ -6692,6 +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"""

convert_hf_to_gguf_update.py

@@ -7,7 +7,6 @@ import pathlib
 import re
 
 import requests
-import sys
 import json
 import shutil
 import argparse
@@ -69,8 +68,7 @@ args = parser.parse_args()
 hf_token = args.hf_token if args.hf_token is not None else hf_token
 
 if hf_token is None:
-    logger.error("HF token is required. Please provide it as an argument or set it in ~/.cache/huggingface/token")
-    sys.exit(1)
+    logger.warning("HF token not found. You can provide it as an argument or set it in ~/.cache/huggingface/token")
 
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
@@ -131,6 +129,7 @@ models = [
     {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
     {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
     {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
+    {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -151,7 +150,7 @@ pre_computed_hashes = [
 
 
 def download_file_with_auth(url, token, save_path):
-    headers = {"Authorization": f"Bearer {token}"}
+    headers = {"Authorization": f"Bearer {token}"} if token else None
     response = sess.get(url, headers=headers)
     response.raise_for_status()
     os.makedirs(os.path.dirname(save_path), exist_ok=True)
@@ -250,10 +249,9 @@ for model in [*pre_computed_hashes, *all_models]:
     else:
         # otherwise, compute the hash of the tokenizer
 
-        # 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
+        # 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}'}...")
@@ -261,9 +259,8 @@ for model in [*pre_computed_hashes, *all_models]:
                 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)

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);
@@ -219,12 +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);
-        //params.sampling.seed++;
+
+        if (sseed < 0) {
+            params.sampling.seed--;
+        }
     }
 
     std::vector<llama_token> tokens_system;

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/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/aclnn_ops.cpp

@@ -1785,8 +1785,27 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
                              bcast_weight_nb[2], bcast_weight_nb[3],
                              bcast_weight_nb[4], bcast_weight_nb[5]};
-    aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_weight_tensor;
+
+    bool weightToNZ = false;
+#ifdef ASCEND_310P
+    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
+#endif
+    if (weightToNZ && is_matmul_weight(weight)) {
+        int64_t acl_stride[2] = {1, transpose_ne[1]};
+
+        // Reverse ne.
+        std::reverse(transpose_ne, transpose_ne + n_dims);
+
+        std::vector<int64_t> storageDims = {transpose_ne[0], transpose_ne[1]};
+
+        acl_weight_tensor = aclCreateTensor(
+            transpose_ne, n_dims, ggml_cann_type_mapping(weight->type), acl_stride,
+            0, ACL_FORMAT_FRACTAL_NZ, storageDims.data(), 2, weight->data);
+    } else {
+        acl_weight_tensor =
+            ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_ND);
+    }
     aclTensor* acl_dst =
         ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
 

ggml/src/ggml-cann/aclnn_ops.h

@@ -23,6 +23,7 @@
 #ifndef CANN_ACLNN_OPS
 #define CANN_ACLNN_OPS
 
+#include <unordered_set>
 #include <functional>
 #include <aclnnop/aclnn_abs.h>
 #include <aclnnop/aclnn_neg.h>
@@ -1020,6 +1021,37 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
  */
 void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
 
+/**
+ * @brief   Check whether a tensor is a weight tensor for matrix multiplication.
+ *
+ * @details Checks whether the given tensor serves as weight parameters in matrix multiplication operations,
+ *          typically within neural network layers. The function maintains a static set of canonical weight
+ *          naming suffixes from Transformer-based architectures. Uses substring matching to identify weight
+ *          tensors even with hierarchical naming patterns.
+ *
+ * @param tensor Pointer to the target ggml_tensor object (const-qualified).
+ */
+static bool is_matmul_weight(const ggml_tensor* tensor) {
+    std::string name = ggml_get_name(tensor);
+    static const std::unordered_set<std::string> weight_suffixes{
+        "output.weight",
+        "attn_q.weight",
+        "attn_k.weight",
+        "attn_v.weight",
+        "attn_output.weight",
+        "ffn_gate.weight",
+        "ffn_up.weight",
+        "ffn_down.weight"
+    };
+
+    for (const auto& suffix : weight_suffixes) {
+        if (name.find(suffix) != std::string::npos) {
+            return true;
+        }
+    }
+    return false;
+}
+
 /**
  * @brief Applies a element-wise operation to two input tensors using the CANN
  * backend.

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

@@ -24,6 +24,7 @@
 
 #include <acl/acl.h>
 #include <stdarg.h>
+#include <aclnnop/aclnn_trans_matmul_weight.h>
 
 #include <cmath>
 #include <cstdio>
@@ -1115,6 +1116,63 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(
     return GGML_STATUS_SUCCESS;
 }
 
+static int CreateAclTensorWeight(const void *hostData, const std::vector<int64_t> &shape, void **deviceAddr,
+                      aclDataType dataType, aclTensor **tensor)
+{
+    uint64_t size = 1;
+    for (auto i : shape) {
+        size *= i;
+    }
+
+    const aclIntArray *mat2Size = aclCreateIntArray(shape.data(), shape.size());
+    ACL_CHECK(aclnnCalculateMatmulWeightSizeV2(mat2Size, dataType, &size));
+
+    size *= sizeof(int16_t);
+
+    ACL_CHECK(aclrtMalloc(deviceAddr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+    aclrtMemcpy(*deviceAddr, size, hostData, size, ACL_MEMCPY_HOST_TO_DEVICE);
+
+    std::vector<int64_t> strides(shape.size(), 1);
+    for (int64_t i = shape.size() - 2; i >= 0; i--) {
+        strides[i] = shape[i + 1] * strides[i + 1];
+    }
+
+    *tensor = aclCreateTensor(shape.data(), shape.size(), dataType, strides.data(), 0, aclFormat::ACL_FORMAT_ND,
+                              shape.data(), shape.size(), *deviceAddr);
+    return 0;
+}
+
+static void weight_format_to_nz(ggml_tensor *tensor, const void *data, size_t offset) {
+    aclrtStream stream;
+    ACL_CHECK(aclrtCreateStream(&stream));
+
+    std::vector<int64_t> weightTransposedShape = {tensor->ne[1], tensor->ne[0]};
+    void *weightTransposedDeviceAddr = nullptr;
+    aclTensor *weightTransposed = nullptr;
+    CreateAclTensorWeight(data, weightTransposedShape, &weightTransposedDeviceAddr,
+                          ggml_cann_type_mapping(tensor->type), &weightTransposed);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor *executor;
+    void *workspaceAddr = nullptr;
+
+    // TransMatmulWeight
+    ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed, &workspaceSize, &executor));
+    std::unique_ptr<void, aclError (*)(void *)> workspaceAddrPtrTrans(nullptr, aclrtFree);
+    if (workspaceSize > 0) {
+        ACL_CHECK(aclrtMalloc(&workspaceAddr, workspaceSize, ACL_MEM_MALLOC_HUGE_FIRST));
+        workspaceAddrPtrTrans.reset(workspaceAddr);
+    }
+    ACL_CHECK(aclnnTransMatmulWeight(workspaceAddr, workspaceSize, executor, stream));
+
+    size_t size = ggml_nelements(tensor) * ggml_element_size(tensor);
+
+    aclrtMemcpy((char *)tensor->data + offset, size,
+                weightTransposedDeviceAddr, size, ACL_MEMCPY_HOST_TO_DEVICE);
+    ACL_CHECK(aclDestroyTensor(weightTransposed));
+    aclrtFree(weightTransposedDeviceAddr);
+}
+
 // TODO: need handle tensor which has paddings.
 /**
  * @brief Set tensor data in a CANN buffer.
@@ -1139,9 +1197,16 @@ static void ggml_backend_cann_buffer_set_tensor(
     // For acl, synchronous functions use this default stream.
     // Why aclrtSynchronizeDevice?
 
+    bool weightToNZ = false;
+#ifdef ASCEND_310P
+    weightToNZ = (getenv("GGML_CANN_WEIGHT_NZ") != nullptr);
+#endif
     if (!need_transform(tensor->type)) {
         ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
                               ACL_MEMCPY_HOST_TO_DEVICE));
+        if (weightToNZ && is_matmul_weight((const ggml_tensor*)tensor)) {
+            weight_format_to_nz(tensor, data, offset);
+        }
     } else {
         void *transform_buffer = malloc(size);
         ggml_backend_cann_transform(tensor, data, transform_buffer);

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/arch/loongarch/quants.c

@@ -544,7 +544,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
         max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
         __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x1 ));
         const float max_scalar = ((v4f32)max4)[0];
 
         // Quantize these floats

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-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/convert.cu

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

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

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

ggml/src/ggml-cuda/cpy.cu

@@ -1,51 +1,17 @@
 #include "cpy.cuh"
 #include "dequantize.cuh"
+#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,
-                                   const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
-                                   const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
+static __global__ void cpy_flt(const char * cx, char * cdst_direct, const int ne,
+                               const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                               const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                               const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) {
     const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
 
     if (i >= ne) {
@@ -71,29 +37,6 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const in
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
-static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q8_0 * dsti = (block_q8_0 *) cdsti;
-
-    float amax = 0.0f; // absolute max
-
-    for (int j = 0; j < QK8_0; j++) {
-        const float v = xi[j];
-        amax = fmaxf(amax, fabsf(v));
-    }
-
-    const float d = amax / ((1 << 7) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK8_0; ++j) {
-        const float x0 = xi[j]*id;
-
-        dsti->qs[j] = roundf(x0);
-    }
-}
-
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
 
@@ -106,139 +49,6 @@ static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     }
 }
 
-static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_0 * dsti = (block_q4_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_0; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -8;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK4_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK4_0/2 + j]*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_1 * dsti = (block_q4_1 *) cdsti;
-
-    float vmin = FLT_MAX;
-    float vmax = -FLT_MAX;
-
-    for (int j = 0; j < QK4_1; ++j) {
-        const float v = xi[j];
-
-        if (v < vmin) vmin = v;
-        if (v > vmax) vmax = v;
-    }
-
-    const float d  = (vmax - vmin) / ((1 << 4) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->dm.x = d;
-    dsti->dm.y = vmin;
-
-    for (int j = 0; j < QK4_1/2; ++j) {
-        const float x0 = (xi[0       + j] - vmin)*id;
-        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q5_0 * dsti = (block_q5_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK5_0; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -16;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK5_0/2 + j]*id;
-
-        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
-        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
-
-        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
-static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q5_1 * dsti = (block_q5_1 *) cdsti;
-
-    float min = xi[0];
-    float max = xi[0];
-
-    for (int j = 1; j < QK5_1; ++j) {
-        const float v = xi[j];
-        min = v < min ? v : min;
-        max = v > max ? v : max;
-    }
-
-    const float d  = (max - min) / 31;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->dm.x = d;
-    dsti->dm.y = min;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_1/2; ++j) {
-        const float x0 = (xi[0       + j] - min)*id;
-        const float x1 = (xi[QK5_1/2 + j] - min)*id;
-
-        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
 template<dequantize_kernel_t dequant, int qk>
 static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
@@ -252,53 +62,6 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     }
 }
 
-static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
-    }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
-}
-
-static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_NL; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    float d = vmax / kvalues_iq4nl[0];
-    const float id = d ? 1.0f/d : 0.0f;
-
-    float sumqx = 0, sumq2 = 0;
-    for (int j = 0; j < QK4_NL/2; ++j) {
-        const float x0 = xi[0        + j]*id;
-        const float x1 = xi[QK4_NL/2 + j]*id;
-        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
-        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
-        dsti->qs[j] = xi0 | (xi1 << 4);
-        const float v0 = kvalues_iq4nl[xi0];
-        const float v1 = kvalues_iq4nl[xi1];
-        const float w0 = xi[0        + j]*xi[0        + j];
-        const float w1 = xi[QK4_NL/2 + j]*xi[QK4_NL/2 + j];
-        sumqx += w0*v0*xi[j] + w1*v1*xi[QK4_NL/2 + j];
-        sumq2 += w0*v0*v0 + w1*v1*v1;
-    }
-
-    dsti->d = sumq2 > 0 ? sumqx/sumq2 : d;
-}
-
 template <cpy_kernel_t cpy_blck, int qk>
 static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@@ -376,43 +139,14 @@ void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_des
 #endif
 }
 
-static void ggml_cpy_f16_f32_cuda(
+template<typename src_t, typename dst_t>
+static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
 
     const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_f32_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_bf16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_bf16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
-static void ggml_cpy_f32_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+    cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
@@ -544,16 +278,6 @@ static void ggml_cpy_f32_iq4_nl_cuda(
         (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
 }
 
-static void ggml_cpy_f16_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
-    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++);
-}
-
 void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1, bool disable_indirection_for_this_node) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
@@ -609,11 +333,11 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
             CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_f32_bf16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<float, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
@@ -640,9 +364,17 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+        ggml_cpy_flt_cuda<half, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_flt_cuda<nv_bfloat16, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_flt_cuda<nv_bfloat16, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
@@ -667,11 +399,11 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         return nullptr;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+        return (void*) cpy_flt<cpy_1_flt<float, float>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_bf16>;
+        return (void*) cpy_flt<cpy_1_flt<float, nv_bfloat16>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+        return (void*) cpy_flt<cpy_1_flt<float, half>>;
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
     } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
@@ -695,9 +427,17 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         return (void*) cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>;
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+        return (void*) cpy_flt<cpy_1_flt<half, half>>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<half, nv_bfloat16>>;
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
-        return (void*) cpy_f32_f16<cpy_1_f16_f32>;
+        return (void*) cpy_flt<cpy_1_flt<half, float>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, half>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, nv_bfloat16>>;
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, float>>;
     } else {
         GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));

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

@@ -745,33 +745,58 @@ void launch_fattn(
     size_t nb23 = V ? V->nb[3] : nb13;
 
     if (need_f16_K && K->type != GGML_TYPE_F16) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(K));
-        K_f16.alloc(ggml_nelements(K));
-        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
-        to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
-        K_data = (char *) K_f16.ptr;
-
         const size_t bs = ggml_blck_size(K->type);
         const size_t ts = ggml_type_size(K->type);
 
-        nb11 = nb11*bs*sizeof(half)/ts;
-        nb12 = nb12*bs*sizeof(half)/ts;
-        nb13 = nb13*bs*sizeof(half)/ts;
+        K_f16.alloc(ggml_nelements(K));
+        if (ggml_is_contiguously_allocated(K)) {
+            to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
+            to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
+
+            nb11 = nb11*bs*sizeof(half)/ts;
+            nb12 = nb12*bs*sizeof(half)/ts;
+            nb13 = nb13*bs*sizeof(half)/ts;
+        } else {
+            GGML_ASSERT(K->nb[0] == ts);
+            to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(K->type);
+            const int64_t s01 = nb11 / ts;
+            const int64_t s02 = nb12 / ts;
+            const int64_t s03 = nb13 / ts;
+            to_fp16(K_data, K_f16.ptr, K->ne[0], K->ne[1], K->ne[2], K->ne[3], s01, s02, s03, main_stream);
+
+            nb11 = K->ne[0] * sizeof(half);
+            nb12 = K->ne[1] * nb11;
+            nb13 = K->ne[2] * nb12;
+        }
+        K_data = (char *) K_f16.ptr;
     }
 
     if (V && need_f16_V && V->type != GGML_TYPE_F16) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(V));
-        V_f16.alloc(ggml_nelements(V));
-        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
-        to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
-        V_data = (char *) V_f16.ptr;
-
         const size_t bs = ggml_blck_size(V->type);
         const size_t ts = ggml_type_size(V->type);
 
-        nb21 = nb21*bs*sizeof(half)/ts;
-        nb22 = nb22*bs*sizeof(half)/ts;
-        nb23 = nb23*bs*sizeof(half)/ts;
+        V_f16.alloc(ggml_nelements(V));
+        if (ggml_is_contiguously_allocated(V)) {
+            to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
+            to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
+            V_data = (char *) V_f16.ptr;
+
+            nb21 = nb21*bs*sizeof(half)/ts;
+            nb22 = nb22*bs*sizeof(half)/ts;
+            nb23 = nb23*bs*sizeof(half)/ts;
+        } else {
+            GGML_ASSERT(V->nb[0] == ts);
+            to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(V->type);
+            const int64_t s01 = nb21 / ts;
+            const int64_t s02 = nb22 / ts;
+            const int64_t s03 = nb23 / ts;
+            to_fp16(V_data, V_f16.ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3], s01, s02, s03, main_stream);
+
+            nb21 = V->ne[0] * sizeof(half);
+            nb22 = V->ne[1] * nb21;
+            nb23 = V->ne[2] * nb22;
+        }
+        V_data = (char *) V_f16.ptr;
     }
 
     int parallel_blocks = 1;

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

@@ -55,6 +55,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <float.h>
+#include <initializer_list>
 #include <limits>
 #include <map>
 #include <memory>
@@ -2590,6 +2591,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
     cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
 
+    const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
+    const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
+
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
@@ -2611,9 +2615,12 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
-            // disable CUDA graphs for batch size > 1 for now.
-            // Changes in batch size or context size can cause changes to the grid size of some kernels.
+        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) && (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true)) {
+            // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
+            // by means of matching node names. See
+            // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
+            // https://github.com/huggingface/transformers/blob/bda75b4011239d065de84aa3e744b67ebfa7b245/src/transformers/models/gemma3n/modeling_gemma3n.py#L1773,
+            // Generally, changes in batch size or context size can cause changes to the grid size of some kernels.
             use_cuda_graph = false;
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
@@ -2759,6 +2766,39 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 }
 #endif
 
+static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
+        return false;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_RMS_NORM && ops.begin()[1] == GGML_OP_MUL) {
+        const ggml_tensor *rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul      = cgraph->nodes[node_idx+1];
+
+        GGML_ASSERT(rms_norm->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(rms_norm->type == GGML_TYPE_F32);
+
+        //rms norm only supports F32
+        if (mul->src[0]->type != GGML_TYPE_F32 ||
+            mul->src[1]->type != GGML_TYPE_F32 ||
+            mul->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        //if rms norm is the B operand, then we don't handle broadcast
+        if (rms_norm == mul->src[1] && !ggml_are_same_shape(mul->src[0], rms_norm->src[1])) {
+            return false;
+        }
+
+        //rms_norm kernel assumes contigous rows
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
     bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
     // flag used to determine whether it is an integrated_gpu
@@ -2768,6 +2808,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
+
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
 
@@ -2775,6 +2816,12 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     continue;
                 }
 
+                static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
+                if (!disable_fusion && ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                    ggml_cuda_op_rms_norm_fused(*cuda_ctx, node, cgraph->nodes[i+1]);
+                    i++;
+                    continue;
+                }
 #ifndef NDEBUG
                 assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
                 for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -3226,8 +3273,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             } break;
         case GGML_OP_SET_ROWS:
             {
-#pragma message("TODO: implement Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
-                return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16) &&
+                return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
+                       op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q5_0 ||
+                       op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_IQ4_NL) &&
                        op->src[0]->type == GGML_TYPE_F32 &&
                        op->src[1]->type == GGML_TYPE_I64;
             } break;
@@ -3235,13 +3283,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1]->type;
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_BF16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                if ((src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_BF16 || src0_type == GGML_TYPE_F16) &&
+                    (src1_type == GGML_TYPE_F32 || src1_type == GGML_TYPE_BF16 || src1_type == GGML_TYPE_F16)
+                ) {
                     return true;
                 }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
@@ -3277,12 +3321,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                     return true;
                 }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
                 if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
                     return true;
                 }
@@ -3363,7 +3401,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return op->src[0]->ne[1] % 128 == 0;
         }
         case GGML_OP_CONT:
-            return op->src[0]->type != GGML_TYPE_BF16;
+            return true;
         case GGML_OP_DIAG_MASK_INF:
             return true;
         case GGML_OP_SOFT_MAX:

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/norm.cu

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

ggml/src/ggml-cuda/norm.cuh

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

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

@@ -1,26 +1,87 @@
 #include "set-rows.cuh"
+#include "cpy-utils.cuh"
 
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
 template<typename src_t, typename dst_t>
-__device__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
-    GGML_UNUSED(src_f);
-    GGML_UNUSED(dst_f);
+__device__ __forceinline__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
+    convert_flt(src_f, dst_f);
 }
 
-template<>
-__device__ __forceinline__ void set_rows_1<float, half>(const float * src_f, half * dst_h) {
-    *dst_h = __float2half(*src_f);
+// Generic quantized set_rows kernel template
+template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+static __global__ void k_set_rows_quant(
+        const float * __restrict__ src0, const int64_t * __restrict__ src1, block_type * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s10, const int64_t s11, const int64_t s12,
+        const int64_t s1, const int64_t s2, const int64_t s3) {
+
+    const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
+    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
+
+    if (i >= ne_total) {
+        return;
+    }
+
+    const int64_t i_base = i * qk;
+    const int64_t i03 = i_base / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i_base - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+
+    const int64_t i12 = i03 % ne12;
+    const int64_t i11 = i02 % ne11;
+    const int64_t i10 = i01;
+
+    const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
+
+    const float * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
+    block_type * dst_row_ptr = dst + (dst_row*s1 + i02*s2 + i03*s3) / sizeof(block_type);
+
+    const float * src_block = src0_row + i00;
+    block_type * dst_block = dst_row_ptr + i00 / qk;
+
+    quantize_func(src_block, dst_block);
 }
 
-template<>
-__device__ __forceinline__ void set_rows_1<float, nv_bfloat16>(const float * src_f, nv_bfloat16 * dst_b) {
-    *dst_b = *src_f;
-}
+// Template dispatch function for quantized set_rows
+template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+static void set_rows_cuda_quant(
+        const float * src0_d, const int64_t * src1_d, block_type * dst_d,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
 
-template<>
-__device__ __forceinline__ void set_rows_1<float, float>(const float * src_f, float * dst_f) {
-    *dst_f = *src_f;
+    GGML_ASSERT(ne00 % qk == 0);
+    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
+    const int num_blocks = (ne_total + CUDA_SET_ROWS_BLOCK_SIZE - 1) / CUDA_SET_ROWS_BLOCK_SIZE;
+    const dim3 block_size(CUDA_SET_ROWS_BLOCK_SIZE);
+    const dim3 grid_size(num_blocks);
+
+    const int64_t s01 = nb01/sizeof(float);
+    const int64_t s02 = nb02/sizeof(float);
+    const int64_t s03 = nb03/sizeof(float);
+    const int64_t s10 = nb10/sizeof(int64_t);
+    const int64_t s11 = nb11/sizeof(int64_t);
+    const int64_t s12 = nb12/sizeof(int64_t);
+    const int64_t s1  = nb1;
+    const int64_t s2  = nb2;
+    const int64_t s3  = nb3;
+
+    if (ne_total > 0) {
+        k_set_rows_quant<block_type, qk, quantize_func><<<grid_size, block_size, 0, stream>>>(
+            src0_d, src1_d, dst_d,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            s01, s02, s03,
+            s10, s11, s12,
+            s1, s2, s3);
+    }
 }
 
 template<typename src_t, typename dst_t>
@@ -145,7 +206,67 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
             nb1, nb2, nb3,
             stream
         );
+    } else if (dst->type == GGML_TYPE_Q4_0) {
+        set_rows_cuda_quant<block_q4_0, QK4_0, quantize_f32_q4_0_block>(
+            src0_d, src1_d, (block_q4_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q4_1) {
+        set_rows_cuda_quant<block_q4_1, QK4_1, quantize_f32_q4_1_block>(
+            src0_d, src1_d, (block_q4_1*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q5_0) {
+        set_rows_cuda_quant<block_q5_0, QK5_0, quantize_f32_q5_0_block>(
+            src0_d, src1_d, (block_q5_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q5_1) {
+        set_rows_cuda_quant<block_q5_1, QK5_1, quantize_f32_q5_1_block>(
+            src0_d, src1_d, (block_q5_1*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q8_0) {
+        set_rows_cuda_quant<block_q8_0, QK8_0, quantize_f32_q8_0_block>(
+            src0_d, src1_d, (block_q8_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_IQ4_NL) {
+        set_rows_cuda_quant<block_iq4_nl, QK4_NL, quantize_f32_iq4_nl_block>(
+            src0_d, src1_d, (block_iq4_nl*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
     } else {
-        GGML_ABORT("unsupported type");
+        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
     }
 }

ggml/src/ggml-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,
@@ -218,6 +259,8 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,
     GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
+    GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL,
+    GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL_ADD,
     GGML_METAL_KERNEL_TYPE_L2_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
     GGML_METAL_KERNEL_TYPE_NORM,
@@ -1135,13 +1178,27 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         // simd_sum and simd_max requires MTLGPUFamilyApple7
 
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                             add,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                         add_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_2,                      add_fuse_2,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_3,                      add_fuse_3,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_4,                      add_fuse_4,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_5,                      add_fuse_5,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_6,                      add_fuse_6,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_7,                      add_fuse_7,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_FUSE_8,                      add_fuse_8,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4,                      add_row_c4,                      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_2,               add_row_c4_fuse_2,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_3,               add_row_c4_fuse_3,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_4,               add_row_c4_fuse_4,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_5,               add_row_c4_fuse_5,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_6,               add_row_c4_fuse_6,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_7,               add_row_c4_fuse_7,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW_C4_FUSE_8,               add_row_c4_fuse_8,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB,                             sub,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW,                         sub_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW_C4,                      sub_row_c4,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                             mul,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                         mul_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW_C4,                      mul_row_c4,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                             div,                             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                         div_row,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW_C4,                      div_row_c4,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32,                      repeat_f32,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16,                      repeat_f16,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32,                      repeat_i32,                      true);
@@ -1206,6 +1263,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,                   set_rows_q5_1,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,                 set_rows_iq4_nl,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                        rms_norm,                        has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL,                    rms_norm_mul,                    has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM_MUL_ADD,                rms_norm_mul_add,                has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_L2_NORM,                         l2_norm,                         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                      group_norm,                      has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                            norm,                            true);
@@ -1893,7 +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,
@@ -1903,7 +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));
 
@@ -1913,7 +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) {
@@ -1924,7 +1986,7 @@ static bool ggml_metal_encode_node(
         case GGML_OP_PERMUTE:
             {
                 // noop -> next node
-            } return true;
+            } return 1;
         default:
             {
             } break;
@@ -1991,6 +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) {
@@ -2062,37 +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,
@@ -2119,12 +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) {
@@ -2132,7 +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)];
                 }
@@ -2257,12 +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);
@@ -2764,7 +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;
@@ -3640,7 +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;
@@ -3656,7 +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
@@ -3664,7 +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
@@ -3673,7 +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;
                     }
 
                     {
@@ -4105,12 +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
 
@@ -4121,23 +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:
             {
@@ -5532,7 +5760,7 @@ static bool ggml_metal_encode_node(
             }
     }
 
-    return true;
+    return n_fuse;
 }
 
 static enum ggml_status ggml_metal_graph_compute(
@@ -6038,20 +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;
@@ -970,46 +994,145 @@ template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat
 
 // assumption: src1 is a row
 // broadcast src1 into src0
-kernel void kernel_add_row(
+template <short F>
+kernel void kernel_add_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] + src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res += src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
-kernel void kernel_sub_row(
+typedef decltype(kernel_add_row_c4_fuse_impl<1>) kernel_add_row_c4_fuse_t;
+
+template [[host_name("kernel_add_row_c4")]]        kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<1>;
+template [[host_name("kernel_add_row_c4_fuse_2")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<2>;
+template [[host_name("kernel_add_row_c4_fuse_3")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<3>;
+template [[host_name("kernel_add_row_c4_fuse_4")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<4>;
+template [[host_name("kernel_add_row_c4_fuse_5")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<5>;
+template [[host_name("kernel_add_row_c4_fuse_6")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<6>;
+template [[host_name("kernel_add_row_c4_fuse_7")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<7>;
+template [[host_name("kernel_add_row_c4_fuse_8")]] kernel kernel_add_row_c4_fuse_t kernel_add_row_c4_fuse_impl<8>;
+
+template <short F>
+kernel void kernel_sub_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] - src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res -= src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
-kernel void kernel_mul_row(
+typedef decltype(kernel_sub_row_c4_fuse_impl<1>) kernel_sub_row_c4_fuse_t;
+
+template [[host_name("kernel_sub_row_c4")]] kernel kernel_sub_row_c4_fuse_t kernel_sub_row_c4_fuse_impl<1>;
+
+template <short F>
+kernel void kernel_mul_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] * src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res *= src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
-kernel void kernel_div_row(
+typedef decltype(kernel_mul_row_c4_fuse_impl<1>) kernel_mul_row_c4_fuse_t;
+
+template [[host_name("kernel_mul_row_c4")]] kernel kernel_mul_row_c4_fuse_t kernel_mul_row_c4_fuse_impl<1>;
+
+template <short F>
+kernel void kernel_div_row_c4_fuse_impl(
         constant ggml_metal_kargs_bin & args,
-        device const float4 * src0,
-        device const float4 * src1,
-        device       float4 * dst,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
         uint tpig[[thread_position_in_grid]]) {
+
     const uint nb = args.ne00/4;
-    dst[tpig] = src0[tpig] / src1[tpig % nb];
+    const uint i  = tpig % nb;
+
+    device const float4 * src0_row = (device const float4 *) (src0);
+    device       float4 *  dst_row = (device       float4 *) (dst);
+
+    device const float4 * src1_row[F];
+    for (short j = 0; j < F; ++j) {
+        src1_row[j] = (device const float4 *) (src1 + args.o1[j]);
+    }
+
+    float4 res = src0_row[tpig];
+
+#pragma unroll(F)
+    for (short j = 0; j < F; ++j) {
+        res /= src1_row[j][i];
+    }
+
+    dst_row[tpig] = res;
 }
 
+typedef decltype(kernel_div_row_c4_fuse_impl<1>) kernel_div_row_c4_fuse_t;
+
+template [[host_name("kernel_div_row_c4")]] kernel kernel_div_row_c4_fuse_t kernel_div_row_c4_fuse_impl<1>;
+
 kernel void kernel_scale(
         device const float * src0,
         device       float * dst,
@@ -2116,26 +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);
@@ -2154,12 +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

@@ -105,6 +105,8 @@ set(GGML_OPENCL_KERNELS
     pad
     repeat
     mul_mat_f16_f32
+    conv2d
+    conv2d_f16_f32
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

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

@@ -390,6 +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;
 
@@ -441,6 +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;
 
@@ -1478,6 +1484,47 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         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
@@ -2361,6 +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:
@@ -4998,6 +5049,82 @@ static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_ten
     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");
+    }
+
+    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);
@@ -6752,6 +6879,12 @@ 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-sycl/ggml-sycl.cpp

@@ -3530,8 +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);
@@ -3575,7 +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 =

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-vulkan/ggml-vulkan.cpp

@@ -328,6 +328,7 @@ struct vk_device_struct {
     uint64_t max_memory_allocation_size;
     uint64_t suballocation_block_size;
     bool fp16;
+    bool bf16;
     bool pipeline_robustness;
     vk::Device device;
     uint32_t vendor_id;
@@ -482,6 +483,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
+    vk_pipeline pipeline_conv2d_f32;
     vk_pipeline pipeline_conv2d_dw_whcn_f32;
     vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
@@ -875,6 +877,38 @@ struct vk_op_rwkv_wkv7_push_constants {
     uint32_t H;
 };
 
+struct vk_op_conv2d_push_constants {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t N;
+
+    uint32_t KW;
+    uint32_t KH;
+    uint32_t W;
+    uint32_t H;
+    uint32_t OW;
+    uint32_t OH;
+
+    uint32_t s0;
+    uint32_t s1;
+    uint32_t p0;
+    uint32_t p1;
+    uint32_t d0;
+    uint32_t d1;
+
+    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;
+};
+
 struct vk_op_conv2d_dw_push_constants {
     uint32_t ne;
     uint32_t batches;
@@ -974,18 +1008,45 @@ private:
 #endif // GGML_VULKAN_MEMORY_DEBUG
 
 class vk_perf_logger {
-public:
+  public:
     void print_timings() {
+        if (timings.empty()) {
+            return;
+        }
+        uint64_t total_all_op_times = 0;
         std::cerr << "----------------\nVulkan Timings:" << std::endl;
-        for (const auto& t : timings) {
-            uint64_t total = 0;
-            for (const auto& time : t.second) {
-                total += time;
+        for (const auto & t : timings) {
+            uint64_t total_op_times = 0;
+            for (const auto & time : t.second) {
+                total_op_times += time;
             }
-            std::cerr << t.first << ": " << t.second.size() << " x " << (total / t.second.size() / 1000.0) << " us" << std::endl;
+            std::cerr << t.first << ": " << t.second.size() << " x " << (total_op_times / t.second.size() / 1000.0)
+                      << " us";
+
+            // If we have as many flops entries as timing entries for the op, then compute and log the flops/S.
+            auto it = flops.find(t.first);
+            if (it != flops.end() && (it->second).size() == t.second.size()) {
+                uint64_t total_op_flops = 0;
+                for (const auto & elem : it->second) {
+                    total_op_flops += elem;
+                }
+                std::cerr << " ("
+                          << (double(total_op_flops) / (1000.0 * 1000.0 * 1000.0)) /
+                                 (double(total_op_times) / (1000.0 * 1000.0 * 1000.0))
+                          << " GFLOPS/s)";
+            }
+
+            total_all_op_times += total_op_times;
+
+            std::cerr << std::endl;
+        }
+
+        if (timings.size() > 0) {
+            std::cerr << "Total time: " << total_all_op_times / 1000.0 << " us." << std::endl;
         }
 
         timings.clear();
+        flops.clear();
     }
 
     void log_timing(const ggml_tensor * node, uint64_t time) {
@@ -994,22 +1055,45 @@ public:
             return;
         }
         if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
-            const uint64_t m = node->src[0]->ne[1];
-            const uint64_t n = node->src[1]->ne[1];
-            const uint64_t k = node->src[1]->ne[0];
-            std::string name = ggml_op_name(node->op);
+            const uint64_t m    = node->src[0]->ne[1];
+            const uint64_t n    = node->src[1]->ne[1];
+            const uint64_t k    = node->src[1]->ne[0];
+            std::string    name = ggml_op_name(node->op);
             if (n == 1) {
                 name += "_VEC m=" + std::to_string(m) + " k=" + std::to_string(k);
             } else {
                 name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
             }
             timings[name].push_back(time);
+            flops[name].push_back(m * n * (k + (k - 1)));
+            return;
+        }
+        if (node->op == GGML_OP_CONV_2D) {
+            std::string   name    = ggml_op_name(node->op);
+            ggml_tensor * knl     = node->src[0];
+            uint64_t      OW      = node->ne[0];
+            uint64_t      OH      = node->ne[1];
+            uint64_t      N       = node->ne[3];
+            uint64_t      Cout    = node->ne[2];
+            uint64_t      KW      = knl->ne[0];
+            uint64_t      KH      = knl->ne[1];
+            uint64_t      Cin     = knl->ne[2];
+            // KxCRS @ CRSxNPQ = KxNPQ -> M=K, K=CRS, N=NPQ
+            uint64_t      size_M  = Cout;
+            uint64_t      size_K  = Cin * KW * KH;
+            uint64_t      size_N  = N * OW * OH;
+            uint64_t      n_flops = size_M * size_N * (size_K + (size_K - 1));
+            name += " M=Cout=" + std::to_string(size_M) + ", K=Cin*KW*KH=" + std::to_string(size_K) +
+                    ", N=N*OW*OH=" + std::to_string(size_N);
+            flops[name].push_back(n_flops);
+            timings[name].push_back(time);
             return;
         }
         timings[ggml_op_name(node->op)].push_back(time);
     }
-private:
+  private:
     std::map<std::string, std::vector<uint64_t>> timings;
+    std::map<std::string, std::vector<uint64_t>> flops;
 };
 
 struct ggml_backend_vk_context {
@@ -2112,6 +2196,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             }
             compile_count++;
         }
+
         compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), spv_size, spv_data, entrypoint,
                                       parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
     };
@@ -2961,6 +3046,42 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    // conv2d
+    uint32_t conv2d_WG_SIZE  = 256;
+    uint32_t conv2d_BS_K     = 128;
+    uint32_t conv2d_BS_CRS   = 16;
+    uint32_t use_collectives = 0;  // Enables subgroup ops for preventing the re-calculation of indices.
+    if (device->subgroup_shuffle &&
+        device->vendor_id != VK_VENDOR_ID_INTEL) {  // Do not enable collectives on Intel, see PR 14316
+        use_collectives = 1;
+        conv2d_BS_CRS   = std::min(
+            device->subgroup_size,
+            conv2d_BS_CRS);  // CRS block size should be capped at sugroup size for correctness when shuffle is used.
+    }
+    uint32_t conv2d_BS_NPQ = 128;
+    uint32_t conv2d_TS_K   = 8;
+    uint32_t conv2d_shmem_req =
+        (conv2d_BS_K * (conv2d_BS_CRS + 1) + conv2d_BS_CRS * (conv2d_BS_NPQ + 1)) * sizeof(float);
+    if (device->properties.limits.maxComputeSharedMemorySize < conv2d_shmem_req) {
+        conv2d_BS_CRS = 8;
+        if (use_collectives) {
+            conv2d_BS_CRS = std::min(device->subgroup_size, conv2d_BS_CRS);
+        }
+    }
+
+    if (use_collectives) {
+        ggml_vk_create_pipeline(
+            device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
+            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
+            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true);
+    } else {
+        ggml_vk_create_pipeline(
+            device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
+            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
+            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true,
+            false);
+    }
+
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f32, "conv2d_dw_cwhn_f32", conv2d_dw_cwhn_f32_len, conv2d_dw_cwhn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
 
@@ -3273,6 +3394,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->fp16 = device->fp16 && vk12_features.shaderFloat16;
 
+#if defined(VK_KHR_shader_bfloat16)
+        device->bf16 = bfloat16_support && bfloat16_features.shaderBFloat16Type;
+#else
+        device->bf16 = false;
+#endif
+
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
         if (device->subgroup_size_control) {
@@ -3615,6 +3742,7 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     bool coopmat_support = false;
     bool coopmat2_support = false;
     bool integer_dot_product = false;
+    bool bfloat16_support = false;
 
     for (auto properties : ext_props) {
         if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
@@ -3635,6 +3763,11 @@ static void ggml_vk_print_gpu_info(size_t idx) {
         } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0 &&
                     !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) {
             integer_dot_product = true;
+#endif
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 &&
+                    !getenv("GGML_VK_DISABLE_BFLOAT16")) {
+            bfloat16_support = true;
 #endif
         }
     }
@@ -3701,10 +3834,25 @@ static void ggml_vk_print_gpu_info(size_t idx) {
         last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_features;
     }
 
+#if defined(VK_KHR_shader_bfloat16)
+    VkPhysicalDeviceShaderBfloat16FeaturesKHR bfloat16_features {};
+    bfloat16_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR;
+    if (bfloat16_support) {
+        last_struct->pNext = (VkBaseOutStructure *)&bfloat16_features;
+        last_struct = (VkBaseOutStructure *)&bfloat16_features;
+    }
+#endif
+
     vkGetPhysicalDeviceFeatures2(physical_device, &device_features2);
 
     fp16 = fp16 && vk12_features.shaderFloat16;
 
+#if defined(VK_KHR_shader_bfloat16)
+    bool bf16 = bfloat16_support && bfloat16_features.shaderBFloat16Type;
+#else
+    bool bf16 = false;
+#endif
+
     uint32_t default_subgroup_size = get_subgroup_size("", device_architecture);
     const size_t subgroup_size = (default_subgroup_size != 0) ? default_subgroup_size : subgroup_props.subgroupSize;
     const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
@@ -3722,8 +3870,8 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none";
 
     std::string device_name = props2.properties.deviceName.data();
-    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | shared memory: %d | int dot: %d | matrix cores: %s\n",
-              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size,
+    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | bf16: %d | warp size: %zu | shared memory: %d | int dot: %d | matrix cores: %s\n",
+              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, bf16, subgroup_size,
               props2.properties.limits.maxComputeSharedMemorySize, integer_dot_product, matrix_cores.c_str());
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
@@ -6809,6 +6957,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_leaky_relu_f32;
         }
         return nullptr;
+    case GGML_OP_CONV_2D:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
+            ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+            return ctx->device->pipeline_conv2d_f32;
+        }
+        return nullptr;
     case GGML_OP_CONV_2D_DW:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             if (ggml_is_contiguous(src1)) {
@@ -7131,6 +7285,31 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             const uint32_t OW = dst->ne[0];
             elements = { N * OC * OH * OW, 1, 1};
         } break;
+    case GGML_OP_CONV_2D:
+        {
+            // src0 - kernel:   [KW, KH, Cin, Cout]
+            // src1 - input:    [W, H, Cin, N]
+            // dst - result:    [OW, OH, Cout, N]
+
+            // Copied from ggml.c: int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d)
+            auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+                return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
+            };
+            // parallelize in {OW/BS_K, OH/BS_NPQ, 1}
+            int64_t W    = src1->ne[0];
+            int64_t H    = src1->ne[1];
+            int64_t KW   = src0->ne[0];
+            int64_t KH   = src0->ne[1];
+            int64_t Cout = src0->ne[3];
+            int64_t N    = src1->ne[3];
+            int64_t OH   = calc_conv_output_size(H, KH, dst->op_params[1], dst->op_params[3], dst->op_params[5]);
+            int64_t OW   = calc_conv_output_size(W, KW, dst->op_params[0], dst->op_params[2], dst->op_params[4]);
+            int64_t NPQ  = N * OW * OH;
+
+            // Tile output matrix to (K/NB_K, NPQ/NB_NPQ, 1) workgroups
+            elements = { static_cast<uint32_t>(Cout), static_cast<uint32_t>(NPQ), 1 };
+        }
+        break;
     case GGML_OP_ADD:
     case GGML_OP_SUB:
     case GGML_OP_DIV:
@@ -7997,6 +8176,55 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
     }, dryrun);
 }
 
+static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
+                            const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb10 == sizeof(float));
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    vk_op_conv2d_push_constants p{};
+    p.Cout = static_cast<uint32_t>(ne03);
+    p.Cin  = static_cast<uint32_t>(ne02);
+    p.N    = static_cast<uint32_t>(ne13);
+
+    p.KW = static_cast<uint32_t>(ne00);
+    p.KH = static_cast<uint32_t>(ne01);
+    p.W  = static_cast<uint32_t>(ne10);
+    p.H  = static_cast<uint32_t>(ne11);
+    p.OW = static_cast<uint32_t>(ne0);
+    p.OH = static_cast<uint32_t>(ne1);
+
+    p.s0 = static_cast<uint32_t>(dst->op_params[0]);
+    p.s1 = static_cast<uint32_t>(dst->op_params[1]);
+    p.p0 = static_cast<uint32_t>(dst->op_params[2]);
+    p.p1 = static_cast<uint32_t>(dst->op_params[3]);
+    p.d0 = static_cast<uint32_t>(dst->op_params[4]);
+    p.d1 = static_cast<uint32_t>(dst->op_params[5]);
+
+    p.nb01 = static_cast<uint32_t>(nb01 / nb00);
+    p.nb02 = static_cast<uint32_t>(nb02 / nb00);
+    p.nb03 = static_cast<uint32_t>(nb03 / nb00);
+
+    p.nb11 = static_cast<uint32_t>(nb11 / nb10);
+    p.nb12 = static_cast<uint32_t>(nb12 / nb10);
+    p.nb13 = static_cast<uint32_t>(nb13 / nb10);
+
+    p.nb1 = static_cast<uint32_t>(nb1 / nb0);
+    p.nb2 = static_cast<uint32_t>(nb2 / nb0);
+    p.nb3 = static_cast<uint32_t>(nb3 / nb0);
+
+    GGML_ASSERT(ne03 == ne2);
+    GGML_ASSERT(ne02 == ne12);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
+}
+
 static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     vk_op_conv2d_dw_push_constants p{};
     p.ne = ggml_nelements(dst);
@@ -9059,6 +9287,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_CONV_TRANSPOSE_1D:
     case GGML_OP_POOL_2D:
+    case GGML_OP_CONV_2D:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
     case GGML_OP_RWKV_WKV7:
@@ -9126,6 +9355,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_POOL_2D:
+        case GGML_OP_CONV_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
             {
@@ -9332,6 +9562,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_POOL_2D:
         ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_CONV_2D:
+        ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_CONV_2D_DW:
         ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -9462,6 +9696,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_CONV_TRANSPOSE_1D:
     case GGML_OP_POOL_2D:
+    case GGML_OP_CONV_2D:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
     case GGML_OP_RWKV_WKV7:
@@ -10013,7 +10248,7 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
         }
         // if rms_norm is the B operand, then we don't handle broadcast
         if (rms_norm == mul->src[1] &&
-            mul->src[0]->ne[1] != rms_norm->ne[1]) {
+            !ggml_are_same_shape(mul->src[0], rms_norm)) {
             return false;
         }
         // rms_norm shader assumes contiguous rows
@@ -10043,6 +10278,12 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
         if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
             total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
+        } else if (cgraph->nodes[i]->op == GGML_OP_CONV_2D) {
+            // Return CRSxNPQxsizeof(*) to account as many bytes as mul_mat has in im2col->mul_mat mode.
+            auto CRS_size =
+                cgraph->nodes[i]->src[0]->ne[0] * cgraph->nodes[i]->src[0]->ne[1] * cgraph->nodes[i]->src[0]->ne[2];
+            auto NPQ_size = cgraph->nodes[i]->ne[0] * cgraph->nodes[i]->ne[1] * cgraph->nodes[i]->ne[3];
+            total_mat_mul_bytes += NPQ_size * CRS_size * ggml_type_size(cgraph->nodes[i]->type);
         }
         i += ctx->num_additional_fused_ops;
         ctx->num_additional_fused_ops = 0;
@@ -10619,6 +10860,20 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             return true;
         case GGML_OP_CONV_TRANSPOSE_1D:
             return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+        case GGML_OP_CONV_2D:
+            {
+                // Op is disabled for Apple because it segfaults at pipeline create time on MoltenVK
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                const vk_device& device = ggml_vk_get_device(ctx->device);
+                bool is_Apple = ggml_vk_get_device(ctx->device)->vendor_id == VK_VENDOR_ID_APPLE;
+                // Channel-contiguous format is not supported yet.
+                return (op->src[0]->type == GGML_TYPE_F32 &&
+                    op->src[1]->type == GGML_TYPE_F32 &&
+                    op->type == GGML_TYPE_F32 &&
+                    ggml_is_contiguous(op->src[0]) &&
+                    ggml_is_contiguous(op->src[1]) &&
+                    ggml_is_contiguous(op)) && !is_Apple;
+            }
         default:
             return false;
     }
@@ -11177,6 +11432,14 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         const int32_t p1 = tensor->op_params[6];
 
         tensor_clone = ggml_pool_2d(ggml_ctx, src_clone[0], op, k0, k1, s0, s1, p0, p1);
+    } else if (tensor->op == GGML_OP_CONV_2D) {
+        const int32_t s0 = tensor->op_params[0];
+        const int32_t s1 = tensor->op_params[1];
+        const int32_t p0 = tensor->op_params[2];
+        const int32_t p1 = tensor->op_params[3];
+        const int32_t d0 = tensor->op_params[4];
+        const int32_t d1 = tensor->op_params[5];
+        tensor_clone = ggml_conv_2d(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
     } else if (tensor->op == GGML_OP_LEAKY_RELU) {
         const float * op_params = (const float *)tensor->op_params;
         tensor_clone = ggml_leaky_relu(ggml_ctx, src_clone[0], op_params[0], false);

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/im2col.comp

@@ -40,12 +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;
@@ -76,7 +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/rms_norm.comp

@@ -50,8 +50,14 @@ void main() {
     const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
 
     if (do_multiply) {
-        [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
-            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+        if (ncols > p.ne10) {
+            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+                data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
+            }
+        } else {
+            [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+                data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+            }
         }
     } else {
         [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {

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

@@ -655,6 +655,8 @@ 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"}}));
 
@@ -765,8 +767,8 @@ void write_output_files() {
                 len += "};\n";
             }
         }
-        fprintf(src, data.c_str());
-        fprintf(src, len.c_str());
+        fputs(data.c_str(), src);
+        fputs(len.c_str(), src);
     }
     fclose(hdr);
     fclose(src);

ggml/src/ggml-webgpu/CMakeLists.txt

@@ -0,0 +1,54 @@
+cmake_minimum_required(VERSION 3.13)
+
+find_package(Python3 REQUIRED)
+
+# Shader locations
+set(SHADER_DIR "${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders")
+set(SHADER_OUTPUT_DIR "${CMAKE_CURRENT_BINARY_DIR}/generated")
+set(SHADER_HEADER "${SHADER_OUTPUT_DIR}/ggml-wgsl-shaders.hpp")
+file(MAKE_DIRECTORY ${SHADER_OUTPUT_DIR})
+
+message(STATUS "Shader output dir: ${SHADER_OUTPUT_DIR}")
+
+# Find all WGSL files
+file(GLOB WGSL_SHADER_FILES "${SHADER_DIR}/*.wgsl")
+
+# Generate the header using a Python script
+add_custom_command(
+    OUTPUT ${SHADER_HEADER}
+    COMMAND ${CMAKE_COMMAND} -E echo "Embedding WGSL shaders to ggml-wgsl-shaders.hpp"
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${SHADER_OUTPUT_DIR}
+    COMMAND ${CMAKE_COMMAND} -E env PYTHONIOENCODING=utf-8
+        ${Python3_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
+            --input "${SHADER_DIR}"
+            --output "${SHADER_HEADER}"
+    DEPENDS ${WGSL_SHADER_FILES} ${CMAKE_CURRENT_SOURCE_DIR}/wgsl-shaders/embed_wgsl.py
+    VERBATIM
+)
+
+add_custom_target(generate_shaders DEPENDS ${SHADER_HEADER})
+
+ggml_add_backend_library(ggml-webgpu
+    ggml-webgpu.cpp
+    ${SHADER_HEADER}
+    ../../include/ggml-webgpu.h
+)
+
+add_dependencies(ggml-webgpu generate_shaders)
+
+if(EMSCRIPTEN)
+    set(EMDAWNWEBGPU_DIR "" CACHE PATH "Path to emdawnwebgpu_pkg")
+
+    target_compile_options(ggml-webgpu PRIVATE "--use-port=${EMDAWNWEBGPU_DIR}/emdawnwebgpu.port.py")
+    target_link_options(ggml-webgpu PRIVATE "--use-port=${EMDAWNWEBGPU_DIR}/emdawnwebgpu.port.py")
+else()
+    find_package(Dawn REQUIRED)
+    set(DawnWebGPU_TARGET dawn::webgpu_dawn)
+endif()
+
+if (GGML_WEBGPU_DEBUG)
+    target_compile_definitions(ggml-webgpu PRIVATE GGML_WEBGPU_DEBUG=1)
+endif()
+
+target_include_directories(ggml-webgpu PRIVATE ${SHADER_OUTPUT_DIR})
+target_link_libraries(ggml-webgpu PRIVATE ${DawnWebGPU_TARGET})

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -0,0 +1,907 @@
+#include "ggml-webgpu.h"
+
+#include <webgpu/webgpu_cpp.h>
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-wgsl-shaders.hpp"
+
+#include <cstring>
+#include <iostream>
+#include <mutex>
+#include <vector>
+
+#ifdef GGML_WEBGPU_DEBUG
+#define WEBGPU_LOG_DEBUG(msg) std::cout << msg << std::endl
+#else
+#define WEBGPU_LOG_DEBUG(msg) ((void) 0)
+#endif // GGML_WEBGPU_DEBUG
+
+/* Constants */
+
+#define WEBGPU_MUL_MAT_WG_SIZE 64
+#define WEBGPU_MUL_MAT_PARAMS_SIZE (13 * sizeof(uint32_t)) // M, N, K, batch sizes, broadcasts
+#define WEBGPU_CPY_PARAMS_SIZE (15 * sizeof(uint32_t)) // strides and offsets
+#define WEBGPU_STORAGE_BUF_BINDING_MULT 4 // a storage buffer binding size must be a multiple of 4
+
+/* End Constants */
+
+// This is a "fake" base pointer, since WebGPU buffers do not have pointers to their locations.
+static void * const webgpu_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
+
+// Always returns the base offset of a tensor, regardless of views.
+static uint64_t webgpu_tensor_offset(const ggml_tensor * tensor) {
+    if (tensor->view_src) {
+        return (uint8_t *) tensor->view_src->data - (uint8_t *) webgpu_ptr_base;
+    }
+    return (uint8_t *) tensor->data - (uint8_t *) webgpu_ptr_base;
+}
+
+/* Struct definitions */
+
+// All the base objects needed to run operations on a WebGPU device
+struct webgpu_context_struct {
+    wgpu::Instance instance;
+    wgpu::Adapter adapter;
+    wgpu::Device device;
+    wgpu::Queue queue;
+    wgpu::Limits limits;
+    wgpu::SupportedFeatures features;
+
+    std::mutex mutex;
+    bool device_initialized = false;
+
+    // pipelines and parameter buffers
+    // TODO: reuse params buffers for different pipelines when possible
+    wgpu::ComputePipeline memset_pipeline;
+    wgpu::Buffer memset_params_dev_buf;
+    wgpu::Buffer memset_params_host_buf;
+    wgpu::ComputePipeline mul_mat_pipeline;
+    wgpu::Buffer mul_mat_params_dev_buf;
+    wgpu::Buffer mul_mat_params_host_buf;
+    wgpu::ComputePipeline cpy_pipeline;
+    wgpu::Buffer cpy_params_dev_buf;
+    wgpu::Buffer cpy_params_host_buf;
+
+    size_t memset_bytes_per_thread;
+
+    // Staging buffer for reading data from the GPU
+    wgpu::Buffer get_tensor_staging_buf;
+};
+
+typedef std::shared_ptr<webgpu_context_struct> webgpu_context;
+
+struct ggml_backend_webgpu_reg_context {
+    webgpu_context webgpu_ctx;
+
+    size_t device_count;
+    const char * name;
+};
+
+struct ggml_backend_webgpu_device_context {
+    webgpu_context webgpu_ctx;
+
+    std::string device_name;
+    std::string device_desc;
+};
+
+struct ggml_backend_webgpu_context {
+    webgpu_context webgpu_ctx;
+
+    std::string name;
+};
+
+struct ggml_backend_webgpu_buffer_context {
+    webgpu_context webgpu_ctx;
+
+    wgpu::Buffer buffer;
+
+    ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf) :
+        webgpu_ctx(ctx), buffer(buf) {
+    }
+};
+
+/* End struct definitions */
+
+/* WebGPU object initializations */
+
+static void ggml_webgpu_create_pipeline(wgpu::Device &device, wgpu::ComputePipeline &pipeline, const char * shader_code, const char * label, const std::vector<wgpu::ConstantEntry> &constants = {}) {
+    WEBGPU_LOG_DEBUG("ggml_webgpu_create_pipeline()");
+    wgpu::ShaderSourceWGSL shader_source;
+    shader_source.code = shader_code;
+    wgpu::ShaderModuleDescriptor shader_desc;
+    shader_desc.nextInChain = &shader_source;
+    wgpu::ShaderModule shader_module = device.CreateShaderModule(&shader_desc);
+
+    wgpu::ComputePipelineDescriptor pipeline_desc;
+    pipeline_desc.label = label;
+    pipeline_desc.compute.module = shader_module;
+    pipeline_desc.compute.entryPoint = "main"; // Entry point in the WGSL code
+    pipeline_desc.layout = nullptr; // nullptr means auto layout
+    if (constants.size() > 0) {
+        pipeline_desc.compute.constants = constants.data();
+        pipeline_desc.compute.constantCount = constants.size();
+    }
+    pipeline = device.CreateComputePipeline(&pipeline_desc);
+}
+
+static void ggml_webgpu_create_buffer(wgpu::Device &device, wgpu::Buffer &buffer, size_t size, wgpu::BufferUsage usage, const char* label) {
+    WEBGPU_LOG_DEBUG("ggml_webgpu_create_buffer()");
+
+    wgpu::BufferDescriptor buffer_desc;
+    buffer_desc.size = size;
+    buffer_desc.usage = usage;
+    buffer_desc.label = label;
+    buffer_desc.mappedAtCreation = false;
+    // TODO: error handling
+    buffer = device.CreateBuffer(&buffer_desc);
+}
+
+/** End WebGPU object initializations */
+
+/** WebGPU Actions */
+
+static void ggml_backend_webgpu_map_buffer(webgpu_context ctx, wgpu::Buffer buffer, wgpu::MapMode mode, size_t offset, size_t size) {
+    ctx->instance.WaitAny(buffer.MapAsync(
+        mode, offset, size, wgpu::CallbackMode::WaitAnyOnly,
+        [](wgpu::MapAsyncStatus status, wgpu::StringView message) {
+            if (status != wgpu::MapAsyncStatus::Success) {
+                GGML_LOG_ERROR("ggml_webgpu: Failed to map buffer: %s\n", message.data);
+            }
+        }),
+        UINT64_MAX
+    );
+}
+
+static void ggml_backend_webgpu_buffer_memset(webgpu_context ctx, wgpu::Buffer buf, uint32_t value, size_t offset, size_t size) {
+    std::lock_guard<std::mutex> lock(ctx->mutex);
+    wgpu::Device device = ctx->device;
+
+    // map the host parameters buffer
+    ggml_backend_webgpu_map_buffer(ctx, ctx->memset_params_host_buf, wgpu::MapMode::Write, 0, ctx->memset_params_host_buf.GetSize());
+    uint32_t * params = (uint32_t *) ctx->memset_params_host_buf.GetMappedRange();
+
+    params[0] = (uint32_t)offset;
+    params[1] = (uint32_t)size;
+    params[2] = value;
+    ctx->memset_params_host_buf.Unmap();
+
+    wgpu::BindGroupEntry entries[2];
+    entries[0].binding = 0; // binding for the buffer to memset
+    entries[0].buffer = buf;
+    entries[0].offset = 0;
+    entries[0].size = buf.GetSize();
+    entries[1].binding = 1; // binding for the parameters
+    entries[1].buffer = ctx->memset_params_dev_buf;
+    entries[1].offset = 0;
+    entries[1].size = ctx->memset_params_dev_buf.GetSize();
+
+    wgpu::BindGroupDescriptor bind_group_desc;
+    bind_group_desc.layout = ctx->memset_pipeline.GetBindGroupLayout(0);
+    bind_group_desc.entryCount = 2;
+    bind_group_desc.label = "ggml_memset";
+    bind_group_desc.entries = entries;
+    wgpu::BindGroup bind_group = device.CreateBindGroup(&bind_group_desc);
+
+    wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
+    encoder.CopyBufferToBuffer(
+        ctx->memset_params_host_buf, 0,
+        ctx->memset_params_dev_buf, 0,
+        ctx->memset_params_dev_buf.GetSize()
+    );
+    wgpu::ComputePassEncoder pass = encoder.BeginComputePass();
+    pass.SetPipeline(ctx->memset_pipeline);
+    pass.SetBindGroup(0, bind_group);
+    size_t bytes_per_wg = ctx->limits.maxComputeWorkgroupSizeX * ctx->memset_bytes_per_thread;
+    pass.DispatchWorkgroups(((size + 3) + bytes_per_wg - 1) / bytes_per_wg, 1, 1);
+    pass.End();
+    wgpu::CommandBuffer commands = encoder.Finish();
+
+    ctx->queue.Submit(1, &commands);
+}
+
+static void ggml_backend_webgpu_wait_on_submission(webgpu_context ctx) {
+    // Wait for the queue to finish processing all commands
+    ctx->instance.WaitAny(ctx->queue.OnSubmittedWorkDone(wgpu::CallbackMode::WaitAnyOnly,
+        [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
+            if (status != wgpu::QueueWorkDoneStatus::Success) {
+                GGML_LOG_ERROR("ggml_webgpu: Failed to wait on queue: %s\n", message.data);
+            }
+        }),
+        UINT64_MAX
+    );
+}
+
+/** End WebGPU Actions */
+
+/** GGML Backend Interface */
+
+static const char * ggml_backend_webgpu_name(ggml_backend_t backend) {
+    ggml_backend_webgpu_context * ctx = (ggml_backend_webgpu_context *)backend->context;
+    return ctx->name.c_str();
+}
+
+static void ggml_backend_webgpu_free(ggml_backend_t backend) {
+    ggml_backend_webgpu_context * ctx = (ggml_backend_webgpu_context *)backend->context;
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_free(" << ctx->name << ")");
+
+    // TODO: cleanup
+    GGML_UNUSED(ctx);
+}
+
+// Returns true if node has enqueued work into the queue, false otherwise
+static bool ggml_webgpu_encode_node(webgpu_context ctx, ggml_tensor * node){
+    if (ggml_is_empty(node)) {
+        return false;
+    }
+
+    WEBGPU_LOG_DEBUG("ggml_webgpu_encode_node(" << node << ", " << ggml_op_name(node->op) << ")");
+
+
+    switch (node->op) {
+        // no-ops
+        case GGML_OP_NONE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+            return false;
+
+        case GGML_OP_CPY: {
+            std::lock_guard<std::mutex> lock(ctx->mutex);
+            const ggml_tensor * src = node->src[0];
+            ggml_backend_webgpu_buffer_context * src_ctx = (ggml_backend_webgpu_buffer_context *) src->buffer->context;
+            size_t src_offset = webgpu_tensor_offset(src) + src->view_offs;
+            // assumes power of 2 offset alignment
+            size_t src_misalignment = src_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+            // align to minimum offset alignment
+            src_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+            ggml_backend_webgpu_buffer_context * dst_ctx = (ggml_backend_webgpu_buffer_context *) node->buffer->context;
+            size_t dst_offset = webgpu_tensor_offset(node) + node->view_offs;
+            size_t dst_misalignment = dst_offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+            dst_offset &= ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+
+            wgpu::Device device = ctx->device;
+            ggml_backend_webgpu_map_buffer(ctx, ctx->cpy_params_host_buf,
+                wgpu::MapMode::Write, 0, ctx->cpy_params_host_buf.GetSize());
+            uint32_t * params = (uint32_t *) ctx->cpy_params_host_buf.GetMappedRange();
+            uint32_t ne = (uint32_t)ggml_nelements(node);
+            params[0] = ne;
+            params[1] = src_misalignment/ggml_type_size(src->type);
+            params[2] = dst_misalignment/ggml_type_size(node->type);
+
+            // Convert byte-strides to element-strides
+            params[3] = (uint32_t)src->nb[0]/ggml_type_size(src->type);
+            params[4] = (uint32_t)src->nb[1]/ggml_type_size(src->type);
+            params[5] = (uint32_t)src->nb[2]/ggml_type_size(src->type);
+            params[6] = (uint32_t)src->nb[3]/ggml_type_size(src->type);
+            params[7] = (uint32_t)node->nb[0]/ggml_type_size(node->type);
+            params[8] = (uint32_t)node->nb[1]/ggml_type_size(node->type);
+            params[9] = (uint32_t)node->nb[2]/ggml_type_size(node->type);
+            params[10] = (uint32_t)node->nb[3]/ggml_type_size(node->type);
+            // Logical shape — same for both tensors even if permuted
+            params[11] = (uint32_t)(src->ne[0]);
+            params[12] = (uint32_t)(src->ne[1]);
+            params[13] = (uint32_t)(src->ne[2]);
+            params[14] = (uint32_t)(src->ne[3]);
+
+            ctx->cpy_params_host_buf.Unmap();
+
+            wgpu::BindGroupEntry entries[3];
+            entries[0].binding = 0;
+            entries[0].buffer = src_ctx->buffer;
+            entries[0].offset = src_offset;
+            entries[0].size = (ggml_nbytes(src) + src_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) & ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1);
+
+            entries[1].binding = 1;
+            entries[1].buffer = dst_ctx->buffer;
+            entries[1].offset = dst_offset;
+            entries[1].size = (ggml_nbytes(node) + dst_misalignment + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) & ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1);
+
+            entries[2].binding = 2;
+            entries[2].buffer = ctx->cpy_params_dev_buf;
+            entries[2].offset = 0;
+            entries[2].size = ctx->cpy_params_dev_buf.GetSize();
+
+            wgpu::BindGroupDescriptor bind_group_desc;
+            bind_group_desc.layout = ctx->cpy_pipeline.GetBindGroupLayout(0);
+            bind_group_desc.label = "ggml_op_cpy";
+            bind_group_desc.entryCount = 3;
+            bind_group_desc.entries = entries;
+            wgpu::BindGroup bind_group = device.CreateBindGroup(&bind_group_desc);
+
+            wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
+            encoder.CopyBufferToBuffer(
+                ctx->cpy_params_host_buf, 0,
+                ctx->cpy_params_dev_buf, 0,
+                ctx->cpy_params_dev_buf.GetSize()
+            );
+            wgpu::ComputePassEncoder pass = encoder.BeginComputePass();
+            pass.SetPipeline(ctx->cpy_pipeline);
+            pass.SetBindGroup(0, bind_group);
+            size_t max_wg_size = ctx->limits.maxComputeWorkgroupSizeX;
+            pass.DispatchWorkgroups((ne + max_wg_size - 1) / max_wg_size);
+            pass.End();
+            wgpu::CommandBuffer commands = encoder.Finish();
+
+            // TODO, don't submit here, batch submissions
+            ctx->queue.Submit(1, &commands);
+            // TODO, don't wait on submission here
+            ggml_backend_webgpu_wait_on_submission(ctx);
+            return true;
+        }
+
+        case GGML_OP_MUL_MAT:
+         {
+            const ggml_tensor * src0 = node->src[0];
+            ggml_backend_webgpu_buffer_context * src0_ctx = (ggml_backend_webgpu_buffer_context *) src0->buffer->context;
+            size_t src0_offset = webgpu_tensor_offset(src0) + src0->view_offs;
+            const ggml_tensor * src1 = node->src[1];
+            ggml_backend_webgpu_buffer_context * src1_ctx = (ggml_backend_webgpu_buffer_context *) src1->buffer->context;
+            size_t src1_offset = webgpu_tensor_offset(src1) + src1->view_offs;
+            ggml_backend_webgpu_buffer_context * dst_ctx = (ggml_backend_webgpu_buffer_context *) node->buffer->context;
+
+            size_t dst_offset = webgpu_tensor_offset(node) + node->view_offs;
+
+            wgpu::Device device = ctx->device;
+
+            // map the host parameters buffer
+            ggml_backend_webgpu_map_buffer(ctx, ctx->mul_mat_params_host_buf,
+                wgpu::MapMode::Write, 0, ctx->mul_mat_params_host_buf.GetSize());
+            uint32_t * params = (uint32_t *) ctx->mul_mat_params_host_buf.GetMappedRange();
+
+            params[0] = (uint32_t)node->ne[1]; // number of rows in result (M)
+            params[1] = (uint32_t)node->ne[0]; // number of columns in result (N)
+            params[2] = (uint32_t)src0->ne[0]; // number of columns in src0/src1 (K)
+
+            params[3] = (uint32_t)src0->nb[1]/ggml_type_size(src0->type); // stride (elements) of src0 in dimension 1
+            params[4] = (uint32_t)src1->nb[1]/ggml_type_size(src1->type); // stride (elements) of src1 in dimension 1
+            params[5] = (uint32_t)src0->nb[2]/ggml_type_size(src0->type); // stride (elements) of src0 in dimension 2
+            params[6] = (uint32_t)src1->nb[2]/ggml_type_size(src1->type); // stride (elements) of src1 in dimension 2
+            params[7] = (uint32_t)src0->nb[3]/ggml_type_size(src0->type); // stride (elements) of src0 in dimension 3
+            params[8] = (uint32_t)src1->nb[3]/ggml_type_size(src1->type); // stride (elements) of src1 in dimension 3
+
+            params[9] = (uint32_t)src0->ne[2]; // batch size in dimension 2
+            params[10] = (uint32_t)src0->ne[3]; // batch size in dimension 3
+            params[11] = (uint32_t)(src1->ne[2]/src0->ne[2]); // broadcast in dimension 2
+            params[12] = (uint32_t)(src1->ne[3]/src0->ne[3]); // broadcast in dimension 3
+
+            ctx->mul_mat_params_host_buf.Unmap();
+
+            wgpu::BindGroupEntry entries[4];
+            entries[0].binding = 0;
+            entries[0].buffer = src0_ctx->buffer;
+            entries[0].offset = src0_offset;
+            entries[0].size = ggml_nbytes(src0);
+
+            entries[1].binding = 1;
+            entries[1].buffer = src1_ctx->buffer;
+            entries[1].offset = src1_offset;
+            entries[1].size = ggml_nbytes(src1);
+
+            entries[2].binding = 2;
+            entries[2].buffer = dst_ctx->buffer;
+            entries[2].offset = dst_offset;
+            entries[2].size = ggml_nbytes(node);
+
+            entries[3].binding = 3;
+            entries[3].buffer = ctx->mul_mat_params_dev_buf;
+            entries[3].offset = 0;
+            entries[3].size = ctx->mul_mat_params_dev_buf.GetSize();
+
+            wgpu::BindGroupDescriptor bind_group_desc;
+            bind_group_desc.layout = ctx->mul_mat_pipeline.GetBindGroupLayout(0);
+            bind_group_desc.entryCount = 4;
+            bind_group_desc.label = "ggml_op_mul_mat";
+            bind_group_desc.entries = entries;
+            wgpu::BindGroup bind_group = device.CreateBindGroup(&bind_group_desc);
+
+            wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
+            encoder.CopyBufferToBuffer(
+                ctx->mul_mat_params_host_buf, 0,
+                ctx->mul_mat_params_dev_buf, 0,
+                ctx->mul_mat_params_dev_buf.GetSize()
+            );
+            wgpu::ComputePassEncoder pass = encoder.BeginComputePass();
+            pass.SetPipeline(ctx->mul_mat_pipeline);
+            pass.SetBindGroup(0, bind_group);
+            pass.DispatchWorkgroups((node->ne[0] * node->ne[1] * node->ne[2] * node->ne[3] + WEBGPU_MUL_MAT_WG_SIZE - 1) / WEBGPU_MUL_MAT_WG_SIZE);
+            pass.End();
+            wgpu::CommandBuffer commands = encoder.Finish();
+
+            // TODO, don't submit here, batch submissions
+            ctx->queue.Submit(1, &commands);
+            // TODO, don't wait on submission here
+            ggml_backend_webgpu_wait_on_submission(ctx);
+            return true;
+        }
+
+        default:
+            return false;
+    }
+}
+
+static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_graph_compute(" << cgraph->n_nodes << " nodes)");
+
+    ggml_backend_webgpu_context * backend_ctx = static_cast<ggml_backend_webgpu_context *>(backend->context);
+    webgpu_context ctx = backend_ctx->webgpu_ctx;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_webgpu_encode_node(ctx, cgraph->nodes[i]);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static ggml_backend_i ggml_backend_webgpu_i = {
+    /* .get_name                = */ ggml_backend_webgpu_name,
+    /* .free                    = */ ggml_backend_webgpu_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_webgpu_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+/* End GGML Backend Interface */
+
+/* GGML Backend Buffer Interface */
+
+static void ggml_backend_webgpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_free_buffer()");
+    ggml_backend_webgpu_buffer_context * ctx = static_cast<ggml_backend_webgpu_buffer_context *>(buffer->context);
+    ctx->buffer.Destroy();
+}
+
+// Returns the "fake" base pointer.
+static void * ggml_backend_webgpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    GGML_UNUSED(buffer);
+    return webgpu_ptr_base;
+}
+
+static void ggml_backend_webgpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    if (size == 0) {
+        WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor: size is zero, nothing to do.");
+        return;
+    }
+
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buffer << ", " << tensor << ", " << value << ", " << offset << ", " << size << ")");
+
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
+    // This is a trick to set all bytes of a u32 to the same 1 byte value.
+    uint32_t val32 = (uint32_t)value * 0x01010101;
+    ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, val32, total_offset, size);
+}
+
+static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")");
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
+
+    size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
+
+    webgpu_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size/4)*4);
+
+    if (size % 4 != 0) {
+        // If size is not a multiple of 4, we need to memset the remaining bytes
+        size_t remaining_size = size % 4;
+        // pack the remaining bytes into a uint32_t
+        uint32_t val32 = 0;
+        for (size_t i = 0; i < remaining_size; i++) {
+            ((uint8_t *)&val32)[i] = ((const uint8_t *)data)[size - remaining_size + i];
+        }
+        // memset the remaining bytes
+        ggml_backend_webgpu_buffer_memset(webgpu_ctx, buf_ctx->buffer, val32, total_offset + (size - remaining_size), remaining_size);
+    }
+}
+
+static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")");
+
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
+    wgpu::Device device = webgpu_ctx->device;
+
+    size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
+
+    size_t final_size = size;
+    if (size % 4 != 0) {
+        // If size is not a multiple of 4, we need to round it up to the next multiple of 4
+        final_size = size + (4 - (size % 4));
+    }
+
+    std::lock_guard<std::mutex> lock(webgpu_ctx->mutex);
+
+    if (webgpu_ctx->get_tensor_staging_buf == nullptr ||
+        webgpu_ctx->get_tensor_staging_buf.GetSize() < final_size) {
+        // Create a new staging buffer if it doesn't exist or is too small
+        if (webgpu_ctx->get_tensor_staging_buf) {
+            webgpu_ctx->get_tensor_staging_buf.Destroy();
+        }
+        ggml_webgpu_create_buffer(device, webgpu_ctx->get_tensor_staging_buf, final_size,
+            wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead, "get_tensor_staging_buf");
+    }
+
+    // Copy the data from the buffer to the staging buffer
+    wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
+    encoder.CopyBufferToBuffer(buf_ctx->buffer, total_offset, webgpu_ctx->get_tensor_staging_buf, 0, final_size);
+    wgpu::CommandBuffer commands = encoder.Finish();
+    // Submit the command buffer to the queue
+    webgpu_ctx->queue.Submit(1, &commands);
+
+    // Map the staging buffer to read the data
+    ggml_backend_webgpu_map_buffer(webgpu_ctx, webgpu_ctx->get_tensor_staging_buf, wgpu::MapMode::Read, 0, final_size);
+    // Must specify size here since the staging buffer might be larger than the tensor size
+    const void * mapped_range = webgpu_ctx->get_tensor_staging_buf.GetConstMappedRange(0, final_size);
+
+    // Copy the data from the mapped range to the output buffer
+    std::memcpy(data, mapped_range, size);
+    webgpu_ctx->get_tensor_staging_buf.Unmap();
+}
+
+static void ggml_backend_webgpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_clear(" << buffer << ", " << (uint32_t) value << ")");
+
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, value, 0, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_webgpu_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_webgpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_webgpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // TODO: optional, needed?
+    /* .memset_tensor   = */ ggml_backend_webgpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_webgpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_webgpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL, // TODO: optional, implement this
+    /* .clear           = */ ggml_backend_webgpu_buffer_clear,
+    /* .reset           = */ NULL, // TODO: optional, think it coordinates with .init_tensor
+};
+
+/* End GGML Backend Buffer Interface */
+
+/* GGML Backend Buffer Type Interface */
+
+static const char * ggml_backend_webgpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return ctx->device_name.c_str();
+}
+
+static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer(" << size << ")");
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+
+    wgpu::Buffer buf;
+    ggml_webgpu_create_buffer(ctx->webgpu_ctx->device, buf, size,
+        wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst, "allocated_buffer");
+
+    ggml_backend_webgpu_buffer_context * buf_ctx = new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_webgpu_buffer_interface, buf_ctx, size);
+}
+
+static size_t ggml_backend_webgpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment;
+}
+
+// maxBufferSize might be larger, but you can't bind more than maxStorageBufferBindingSize to a single binding.
+static size_t ggml_backend_webgpu_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return ctx->webgpu_ctx->limits.maxStorageBufferBindingSize;
+}
+
+/* End GGML Backend Buffer Type Interface */
+
+/* GGML Backend Device Interface */
+
+static const char * ggml_backend_webgpu_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    return ctx->device_name.c_str();
+}
+
+static const char * ggml_backend_webgpu_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    return ctx->device_desc.c_str();
+}
+
+static void ggml_backend_webgpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    // TODO: what do we actually want to return here? maxBufferSize might not be the full available memory.
+    *free = ctx->webgpu_ctx->limits.maxBufferSize;
+    *total = ctx->webgpu_ctx->limits.maxBufferSize;
+}
+
+static enum ggml_backend_dev_type ggml_backend_webgpu_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+}
+
+static void ggml_backend_webgpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_webgpu_device_get_name(dev);
+    props->description = ggml_backend_webgpu_device_get_description(dev);
+    props->type        = ggml_backend_webgpu_device_get_type(dev);
+    ggml_backend_webgpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_guid_t ggml_backend_webgpu_guid(void) {
+    static const char * guid_str = "__ggml_webgpu :)";
+    return reinterpret_cast<ggml_guid_t>((void *)guid_str);
+}
+
+static void ggml_webgpu_init_memset_pipeline(webgpu_context webgpu_ctx) {
+    // we use the maximum workgroup size for the memset pipeline
+    size_t max_wg_size = webgpu_ctx->limits.maxComputeWorkgroupSizeX;
+    size_t max_threads = max_wg_size * webgpu_ctx->limits.maxComputeWorkgroupsPerDimension;
+    // Size the bytes_per_thread so that the largest buffer size can be handled
+    webgpu_ctx->memset_bytes_per_thread = (webgpu_ctx->limits.maxStorageBufferBindingSize + max_threads - 1) / max_threads;
+    std::vector<wgpu::ConstantEntry> constants(2);
+    constants[0].key = "wg_size";
+    constants[0].value = max_wg_size;
+    constants[1].key = "bytes_per_thread";
+    constants[1].value = webgpu_ctx->memset_bytes_per_thread;
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->memset_pipeline, wgsl_memset, "memset", constants);
+    ggml_webgpu_create_buffer(webgpu_ctx->device, webgpu_ctx->memset_params_dev_buf,
+        3 * sizeof(uint32_t), // 3 parameters: buffer size, offset, value
+        wgpu::BufferUsage::Uniform | wgpu::BufferUsage::CopyDst, "memset_params_dev_buf");
+    ggml_webgpu_create_buffer(webgpu_ctx->device, webgpu_ctx->memset_params_host_buf,
+        3 * sizeof(uint32_t), wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc, "memset_params_host_buf");
+}
+
+static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context webgpu_ctx) {
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline, wgsl_mul_mat, "mul_mat");
+    ggml_webgpu_create_buffer(webgpu_ctx->device, webgpu_ctx->mul_mat_params_dev_buf, WEBGPU_MUL_MAT_PARAMS_SIZE,
+        wgpu::BufferUsage::Uniform | wgpu::BufferUsage::CopyDst, "mul_mat_params_dev_buf");
+    ggml_webgpu_create_buffer(webgpu_ctx->device, webgpu_ctx->mul_mat_params_host_buf, WEBGPU_MUL_MAT_PARAMS_SIZE,
+        wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc, "mul_mat_params_host_buf");
+}
+
+static void ggml_webgpu_init_cpy_pipeline(webgpu_context webgpu_ctx) {
+    std::vector<wgpu::ConstantEntry> constants(1);
+    constants[0].key = "wg_size";
+    constants[0].value = webgpu_ctx->limits.maxComputeWorkgroupSizeX;
+
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->cpy_pipeline, wgsl_cpy, "cpy", constants);
+    ggml_webgpu_create_buffer(webgpu_ctx->device, webgpu_ctx->cpy_params_dev_buf, WEBGPU_CPY_PARAMS_SIZE,
+        wgpu::BufferUsage::Uniform | wgpu::BufferUsage::CopyDst, "cpy_params_dev_buf");
+    ggml_webgpu_create_buffer(webgpu_ctx->device, webgpu_ctx->cpy_params_host_buf, WEBGPU_CPY_PARAMS_SIZE,
+        wgpu::BufferUsage::MapWrite | wgpu::BufferUsage::CopySrc, "cpy_params_host_buf");
+}
+
+// TODO: Make thread safe if multiple devices are used
+static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_device_init()");
+
+    ggml_backend_webgpu_device_context * dev_ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
+    webgpu_context webgpu_ctx = dev_ctx->webgpu_ctx;
+
+    std::lock_guard<std::mutex> lock(webgpu_ctx->mutex);
+
+    if (!webgpu_ctx->device_initialized) {
+        // Initialize device
+        wgpu::DeviceDescriptor dev_desc;
+        dev_desc.requiredLimits = &webgpu_ctx->limits;
+        dev_desc.requiredFeatures = webgpu_ctx->features.features;
+        dev_desc.requiredFeatureCount = webgpu_ctx->features.featureCount;
+        dev_desc.SetDeviceLostCallback(wgpu::CallbackMode::AllowSpontaneous,
+            [](const wgpu::Device& device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
+                GGML_UNUSED(device);
+                GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
+        });
+        dev_desc.SetUncapturedErrorCallback(
+            [](const wgpu::Device& device, wgpu::ErrorType reason, wgpu::StringView message) {
+                GGML_UNUSED(device);
+                GGML_LOG_ERROR("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason), message.data);
+        });
+        webgpu_ctx->instance.WaitAny(webgpu_ctx->adapter.RequestDevice(&dev_desc, wgpu::CallbackMode::WaitAnyOnly,
+            [webgpu_ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
+                if (status != wgpu::RequestDeviceStatus::Success) {
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", message.data);
+                    return;
+                }
+                webgpu_ctx->device = device;
+            }),
+            UINT64_MAX
+        );
+        GGML_ASSERT(webgpu_ctx->device != nullptr);
+
+        // Initialize (compute) queue
+        webgpu_ctx->queue = webgpu_ctx->device.GetQueue();
+
+        ggml_webgpu_init_memset_pipeline(webgpu_ctx);
+        ggml_webgpu_init_mul_mat_pipeline(webgpu_ctx);
+        ggml_webgpu_init_cpy_pipeline(webgpu_ctx);
+        webgpu_ctx->device_initialized = true;
+    }
+
+    static ggml_backend_webgpu_context backend_ctx;
+    backend_ctx.name = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name;
+    backend_ctx.webgpu_ctx = webgpu_ctx;
+
+    // See GGML Backend Interface section
+    static ggml_backend backend = {
+        /* .guid      = */ ggml_backend_webgpu_guid(),
+        /* .interface = */ ggml_backend_webgpu_i,
+        /* .device    = */ dev,
+        /* .context   = */ &backend_ctx,
+    };
+
+    return &backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggml_backend_dev_t dev) {
+    // See GGML Backend Buffer Type Interface section
+    static struct ggml_backend_buffer_type ggml_backend_webgpu_buffer_type = {
+        /* .iface = */ {
+            /* .get_name         = */ ggml_backend_webgpu_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_webgpu_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_webgpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ ggml_backend_webgpu_buffer_type_get_max_size,
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ NULL, // defaults to false
+        },
+        /* .device  = */ dev,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_webgpu_buffer_type;
+}
+
+static bool ggml_backend_webgpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    GGML_UNUSED(dev);
+    return  buft->iface.get_name == ggml_backend_webgpu_buffer_type_get_name;
+}
+
+static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    GGML_UNUSED(dev);
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+            return true;
+        case GGML_OP_CPY:
+            return op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_MUL_MAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+        default:
+            return false;
+    }
+}
+
+static struct ggml_backend_device_i ggml_backend_webgpu_device_i = {
+    /* .get_name             = */ ggml_backend_webgpu_device_get_name,
+    /* .get_description      = */ ggml_backend_webgpu_device_get_description,
+    /* .get_memory           = */ ggml_backend_webgpu_device_get_memory,
+    /* .get_type             = */ ggml_backend_webgpu_device_get_type,
+    /* .get_props            = */ ggml_backend_webgpu_device_get_props,
+    /* .init_backend         = */ ggml_backend_webgpu_device_init,
+    /* .get_buffer_type      = */ ggml_backend_webgpu_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ NULL,
+    /* .supports_op          = */ ggml_backend_webgpu_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_webgpu_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+/* End GGML Backend Device Interface */
+
+/* GGML Backend Registration Interface */
+
+static const char * ggml_backend_webgpu_reg_get_name(ggml_backend_reg_t reg) {
+    ggml_backend_webgpu_reg_context * ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
+    return ctx->name;
+}
+
+static size_t ggml_backend_webgpu_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_webgpu_reg_context * ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
+    return ctx->device_count;
+}
+
+// TODO: Does this need to be thread safe? Is it only called once?
+// Only one device is supported for now
+static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+    WEBGPU_LOG_DEBUG("ggml_backend_reg_get_device()");
+
+    ggml_backend_webgpu_reg_context * reg_ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
+
+    webgpu_context ctx = reg_ctx->webgpu_ctx;
+
+    wgpu::RequestAdapterOptions options = {};
+    auto callback = [](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char *message, void *userdata) {
+        if (status != wgpu::RequestAdapterStatus::Success) {
+            GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
+            return;
+        }
+        *static_cast<wgpu::Adapter *>(userdata) = adapter;
+    };
+    void *userdata = &ctx->adapter;
+    ctx->instance.WaitAny(ctx->instance.RequestAdapter(&options, wgpu::CallbackMode::WaitAnyOnly, callback, userdata), UINT64_MAX);
+    GGML_ASSERT(ctx->adapter != nullptr);
+
+    ctx->adapter.GetLimits(&ctx->limits);
+    ctx->adapter.GetFeatures(&ctx->features);
+
+    wgpu::AdapterInfo info{};
+    ctx->adapter.GetInfo(&info);
+
+    static ggml_backend_webgpu_device_context device_ctx;
+    device_ctx.webgpu_ctx = ctx;
+    device_ctx.device_name = GGML_WEBGPU_NAME;
+    device_ctx.device_desc = std::string(info.description.data);
+
+    GGML_LOG_INFO("ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | device_desc: %s\n",
+        info.vendorID, info.vendor.data, info.architecture.data, info.deviceID, info.device.data, info.description.data);
+
+    // See GGML Backend Device Interface section
+    static ggml_backend_device device = {
+        /* .iface   = */ ggml_backend_webgpu_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ &device_ctx,
+    };
+    return &device;
+}
+
+
+static const struct ggml_backend_reg_i ggml_backend_webgpu_reg_i = {
+    /* .get_name         = */ ggml_backend_webgpu_reg_get_name,
+    /* .get_device_count = */ ggml_backend_webgpu_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_webgpu_reg_get_device,
+    /* .get_proc_address = */ NULL,
+};
+
+/* End GGML Backend Registration Interface */
+
+// TODO: Does this need to be thread safe? Is it only called once?
+ggml_backend_reg_t ggml_backend_webgpu_reg() {
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_reg()");
+
+    webgpu_context webgpu_ctx = std::make_shared<webgpu_context_struct>();
+    webgpu_ctx->device_initialized = false;
+
+    static ggml_backend_webgpu_reg_context ctx;
+    ctx.webgpu_ctx = webgpu_ctx;
+    ctx.name = GGML_WEBGPU_NAME;
+    ctx.device_count = 1;
+
+    wgpu::InstanceDescriptor instance_descriptor{};
+    std::vector<wgpu::InstanceFeatureName> instance_features = {wgpu::InstanceFeatureName::TimedWaitAny};
+    instance_descriptor.requiredFeatures = instance_features.data();
+    instance_descriptor.requiredFeatureCount = instance_features.size();
+    webgpu_ctx->instance = wgpu::CreateInstance(&instance_descriptor);
+    GGML_ASSERT(webgpu_ctx->instance != nullptr);
+
+    static ggml_backend_reg reg = {
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_webgpu_reg_i,
+        /* .context     = */ &ctx,
+    };
+    return &reg;
+}
+
+ggml_backend_t ggml_backend_webgpu_init(void) {
+    ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_webgpu_reg(), 0);
+
+    return ggml_backend_webgpu_device_init(dev, nullptr);
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_webgpu_reg)

ggml/src/ggml-webgpu/wgsl-shaders/cpy.wgsl

@@ -0,0 +1,60 @@
+enable f16;
+
+@group(0) @binding(0)
+var<storage, read_write> src: array<f32>;
+
+@group(0) @binding(1)
+var<storage, read_write> dst: array<f16>;
+
+struct Params {
+    ne: u32,            // total number of elements
+    offset_src: u32,    // in elements
+    offset_dst: u32,    // in elements
+
+    // Strides (in elements) — may be permuted
+    stride_src0: u32,
+    stride_src1: u32,
+    stride_src2: u32,
+    stride_src3: u32,
+
+    stride_dst0: u32,
+    stride_dst1: u32,
+    stride_dst2: u32,
+    stride_dst3: u32,
+
+    // Logical shape (same for both tensors)
+    ne0: u32,
+    ne1: u32,
+    ne2: u32,
+    ne3: u32,
+};
+
+@group(0) @binding(2)
+var<uniform> params: Params;
+
+override wg_size: u32;
+@compute @workgroup_size(wg_size)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    if (gid.x >= params.ne) {
+        return;
+    }
+
+    var i = gid.x;
+
+    let i3 = i / (params.ne2 * params.ne1 * params.ne0);
+    i = i % (params.ne2 * params.ne1 * params.ne0);
+
+    let i2 = i / (params.ne1 * params.ne0);
+    i = i % (params.ne1 * params.ne0);
+
+    let i1 = i / params.ne0;
+    let i0 = i % params.ne0;
+
+    let src_idx = i0 * params.stride_src0 + i1 * params.stride_src1 +
+                  i2 * params.stride_src2 + i3 * params.stride_src3;
+
+    let dst_idx = i0 * params.stride_dst0 + i1 * params.stride_dst1 +
+                  i2 * params.stride_dst2 + i3 * params.stride_dst3;
+
+    dst[params.offset_dst + dst_idx] = f16(src[params.offset_src + src_idx]);
+}

ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py

@@ -0,0 +1,35 @@
+import os
+import argparse
+
+
+def escape_triple_quotes(wgsl):
+    # Simple defense in case of embedded """
+    return wgsl.replace('"""', '\\"""')
+
+
+def to_cpp_string_literal(varname, content):
+    return f'const char* wgsl_{varname} = R"({content})";\n'
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--input', required=True)
+    parser.add_argument('--output', required=True)
+    args = parser.parse_args()
+
+    with open(args.output, 'w', encoding='utf-8') as out:
+        out.write("// Auto-generated shader embedding \n\n")
+        for fname in sorted(os.listdir(args.input)):
+            if not fname.endswith('.wgsl'):
+                continue
+            shader_path = os.path.join(args.input, fname)
+            varname = os.path.splitext(fname)[0]
+            with open(shader_path, 'r', encoding='utf-8') as f:
+                content = f.read()
+            content = escape_triple_quotes(content)
+            out.write(to_cpp_string_literal(varname, content))
+            out.write('\n')
+
+
+if __name__ == '__main__':
+    main()

ggml/src/ggml-webgpu/wgsl-shaders/memset.wgsl

@@ -0,0 +1,40 @@
+@group(0) @binding(0)
+var<storage, read_write> output_buffer: array<u32>;
+
+struct Params {
+    offset: u32, // in bytes
+    size: u32,   // in bytes
+    value: u32,  // 4 8-bit values, which are either repeating (memset_tensor) or may be separate (cleaning up unaligned set_tensor operations)
+};
+
+@group(0) @binding(1)
+var<uniform> params: Params;
+
+override wg_size: u32;
+override bytes_per_thread: u32;
+
+@compute @workgroup_size(wg_size)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let i = gid.x * bytes_per_thread;
+    let start = params.offset;
+    let end = params.offset + params.size;
+
+    for (var j: u32 = 0u; j < bytes_per_thread; j = j + 1u) {
+        let byte_index = start + i + j;
+        if (byte_index + 4u <= end) {
+            output_buffer[(byte_index >> 2u)] = params.value;
+        } else {
+            // Handle tail (unaligned)
+            for (var k: u32 = 0u; k < 4u; k = k + 1u) {
+                let idx = byte_index + k;
+                if (idx < end) {
+                    let word_idx = idx >> 2u;
+                    let byte_offset = (idx & 3u) * 8u;
+                    let mask = ~(0xffu << byte_offset);
+                    let existing = output_buffer[word_idx];
+                    output_buffer[word_idx] = (existing & mask) | ((params.value & 0xffu) << byte_offset);
+                }
+            }
+        }
+    }
+}

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.wgsl

@@ -0,0 +1,56 @@
+struct MulMatParams {
+    m: u32,
+    n: u32,
+    k: u32,
+    // all strides are in elements
+    stride_01: u32,
+    stride_11: u32,
+    stride_02: u32,
+    stride_12: u32,
+    stride_03: u32,
+    stride_13: u32,
+
+    bs02: u32,
+    bs03: u32,
+    broadcast2: u32,
+    broadcast3: u32
+};
+
+@group(0) @binding(0) var<storage, read_write> src0: array<f32>; // N rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<f32>; // M rows, K columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
+
+@group(0) @binding(3) var<uniform> params: MulMatParams;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
+    let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
+    if (global_id.x >= total) {
+        return;
+    }
+
+    let dst2_stride = params.m * params.n;
+    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
+
+    let dst3_idx = global_id.x / dst3_stride;
+    let src03_idx = dst3_idx / params.broadcast3; // src0 may be broadcast along the third dimension
+    let src13_idx = dst3_idx; // src1 is not broadcast
+    let dst3_rem = global_id.x % dst3_stride;
+
+    let dst2_idx = dst3_rem / dst2_stride;
+    let src02_idx = dst2_idx / params.broadcast2; // src0 may also be broadcast along the second dimension
+    let src12_idx = dst2_idx; // src1 is not broadcast
+
+    let dst2_rem = dst3_rem % dst2_stride;
+
+    let row = dst2_rem / params.n; // output row
+    let col = dst2_rem % params.n; // output column
+
+    var sum = 0.0;
+    for (var i: u32 = 0u; i < params.k; i = i + 1u) {
+        let src0_idx = src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01 + i;
+        let src1_idx = src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11 + i;
+        sum = sum + src0[src0_idx] * src1[src1_idx];
+    }
+    dst[dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
+}

gguf-py/gguf/constants.py

@@ -233,6 +233,11 @@ class Keys:
         TYPE       = "adapter.type"
         LORA_ALPHA = "adapter.lora.alpha"
 
+    class IMatrix:
+        CHUNK_COUNT = "imatrix.chunk_count"
+        CHUNK_SIZE  = "imatrix.chunk_size"
+        DATASETS    = "imatrix.datasets"
+
     class Clip:
         PROJECTOR_TYPE      = "clip.projector_type"
         HAS_VISION_ENCODER  = "clip.has_vision_encoder"
@@ -282,6 +287,7 @@ class Keys:
 class GGUFType:
     MODEL   = "model"
     ADAPTER = "adapter"
+    IMATRIX = "imatrix"
     MMPROJ  = "mmproj" # dummy, unused for now
 
 
@@ -354,6 +360,7 @@ class MODEL_ARCH(IntEnum):
     JAIS             = auto()
     NEMOTRON         = auto()
     EXAONE           = auto()
+    EXAONE4          = auto()
     GRANITE          = auto()
     GRANITE_MOE      = auto()
     GRANITE_HYBRID   = auto()
@@ -364,6 +371,7 @@ class MODEL_ARCH(IntEnum):
     DOTS1            = auto()
     ARCEE            = auto()
     ERNIE4_5         = auto()
+    ERNIE4_5_MOE     = auto()
     HUNYUAN_MOE      = auto()
     SMOLLM3          = auto()
     LFM2             = auto()
@@ -670,6 +678,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.JAIS:             "jais",
     MODEL_ARCH.NEMOTRON:         "nemotron",
     MODEL_ARCH.EXAONE:           "exaone",
+    MODEL_ARCH.EXAONE4:          "exaone4",
     MODEL_ARCH.GRANITE:          "granite",
     MODEL_ARCH.GRANITE_MOE:      "granitemoe",
     MODEL_ARCH.GRANITE_HYBRID:   "granitehybrid",
@@ -680,6 +689,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.DOTS1:            "dots1",
     MODEL_ARCH.ARCEE:            "arcee",
     MODEL_ARCH.ERNIE4_5:         "ernie4_5",
+    MODEL_ARCH.ERNIE4_5_MOE:     "ernie4_5-moe",
     MODEL_ARCH.FALCON_H1:        "falcon-h1",
     MODEL_ARCH.HUNYUAN_MOE:      "hunyuan-moe",
     MODEL_ARCH.SMOLLM3:          "smollm3",
@@ -2022,6 +2032,28 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FFN_EXP_PROBS_B,
     ],
+    MODEL_ARCH.ERNIE4_5_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+    ],
     MODEL_ARCH.PLM: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT,
@@ -2173,6 +2205,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.EXAONE4: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_POST_NORM,
+    ],
     MODEL_ARCH.GRANITE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/metadata.py

@@ -144,6 +144,10 @@ class Metadata:
         # Quick hack to fix the Norway problem
         # https://hitchdev.com/strictyaml/why/implicit-typing-removed/
         yaml_content = yaml_content.replace("- no\n", "- \"no\"\n")
+        # yaml should use 2 spaces insted of tab
+        # this issue has came up with the Qwen/Qwen3-235B-A22B-Instruct-2507 model card
+        #    (I've also sent a pr tp fix the modelcard too)
+        yaml_content = yaml_content.replace("\t", "  ")
 
         if yaml_content:
             data = yaml.safe_load(yaml_content)

gguf-py/gguf/tensor_mapping.py

@@ -324,7 +324,8 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
-            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
+            "model.layers.{bid}.mlp.gate.e_score_correction",               # deepseek-v3 dots1
+            "model.layers.{bid}.mlp.moe_statics.e_score_correction",        # ernie4.5-moe
         ),
 
         # Feed-forward up
@@ -364,13 +365,13 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
-            "layers.{bid}.feed_forward.experts.w3",           # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear_v",   # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.v1",    # dbrx
-            "model.layers.{bid}.mlp.experts.up_proj",         # qwen2moe olmoe (merged)
-            "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
-            "model.layers.{bid}.feed_forward.experts.up_proj", # llama4
-            "encoder.layers.{bid}.mlp.experts.mlp.w1",        # nomic-bert-moe
+            "layers.{bid}.feed_forward.experts.w3",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_v",         # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.v1",          # dbrx
+            "model.layers.{bid}.mlp.experts.up_proj",               # qwen2moe olmoe (merged) ernie4.5-moe
+            "model.layers.{bid}.block_sparse_moe.experts.w3",       # phimoe (merged)
+            "model.layers.{bid}.feed_forward.experts.up_proj",      # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w1",              # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_UP_SHEXP: (
@@ -403,12 +404,12 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
-            "layers.{bid}.feed_forward.experts.w1",              # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear",        # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.w1",       # dbrx
-            "model.layers.{bid}.mlp.experts.gate_proj",          # qwen2moe olmoe (merged)
-            "model.layers.{bid}.block_sparse_moe.experts.w1",    # phimoe (merged)
-            "model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
+            "layers.{bid}.feed_forward.experts.w1",                     # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear",               # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w1",              # dbrx
+            "model.layers.{bid}.mlp.experts.gate_proj",                 # qwen2moe olmoe (merged) ernie4.5-moe
+            "model.layers.{bid}.block_sparse_moe.experts.w1",           # phimoe (merged)
+            "model.layers.{bid}.feed_forward.experts.gate_proj",        # llama4
         ),
 
         MODEL_TENSOR.FFN_GATE_SHEXP: (
@@ -450,14 +451,14 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
-            "layers.{bid}.feed_forward.experts.w2",              # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear_1",      # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.w2",       # dbrx
-            "model.layers.{bid}.mlp.experts.down_proj",          # qwen2moe olmoe (merged)
-            "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
-            "model.layers.{bid}.block_sparse_moe.experts.w2",    # phimoe (merged)
-            "model.layers.{bid}.feed_forward.experts.down_proj", # llama4
-            "encoder.layers.{bid}.mlp.experts.mlp.w2",           # nomic-bert-moe
+            "layers.{bid}.feed_forward.experts.w2",                 # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear_1",         # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w2",          # dbrx
+            "model.layers.{bid}.mlp.experts.down_proj",             # qwen2moe olmoe (merged) ernie4.5-moe
+            "model.layers.{bid}.block_sparse_moe.output_linear",    # granitemoe
+            "model.layers.{bid}.block_sparse_moe.experts.w2",       # phimoe (merged)
+            "model.layers.{bid}.feed_forward.experts.down_proj",    # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w2",              # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_DOWN_SHEXP: (

include/llama.h

@@ -1394,6 +1394,7 @@ extern "C" {
 
         int32_t n_p_eval;
         int32_t n_eval;
+        int32_t n_reused; // number of times a ggml compute graph had been reused
     };
 
     struct llama_perf_sampler_data {

scripts/sync-ggml.last

@@ -1 +1 @@
-d62df60a07ba3deeb85e5cfc9b1ee07645ff35e2
+3323219cd3cc050e5c7133cd4fc1e50d1f590faf

src/llama-arch.cpp

@@ -68,6 +68,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_JAIS,             "jais"             },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_EXAONE,           "exaone"           },
+    { LLM_ARCH_EXAONE4,          "exaone4"          },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
     { LLM_ARCH_RWKV6QWEN2,       "rwkv6qwen2"       },
     { LLM_ARCH_RWKV7,            "rwkv7"            },
@@ -82,6 +83,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
+    { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },
     { LLM_ARCH_SMOLLM3,          "smollm3"          },
     { LLM_ARCH_LFM2,             "lfm2"             },
@@ -1509,6 +1511,26 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_EXAONE4,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        }
+    },
     {
         LLM_ARCH_RWKV6,
         {
@@ -1825,6 +1847,31 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_ERNIE4_5_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        },
+    },
     {
         LLM_ARCH_HUNYUAN_MOE,
         {

src/llama-arch.h

@@ -72,6 +72,7 @@ enum llm_arch {
     LLM_ARCH_JAIS,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_EXAONE,
+    LLM_ARCH_EXAONE4,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
     LLM_ARCH_RWKV7,
@@ -86,6 +87,7 @@ enum llm_arch {
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
     LLM_ARCH_ERNIE4_5,
+    LLM_ARCH_ERNIE4_5_MOE,
     LLM_ARCH_HUNYUAN_MOE,
     LLM_ARCH_SMOLLM3,
     LLM_ARCH_LFM2,

src/llama-batch.cpp

@@ -157,6 +157,8 @@ bool llama_batch_allocr::init(
         n_outputs += batch.logits[i] != 0;
     }
 
+    has_cpl = false;
+
     // determine coupled sequences
     // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
@@ -208,7 +210,7 @@ bool llama_batch_allocr::init(
         LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__);
 
         llama_ubatch ubatch {
-            /*.equal_seqs   =*/ false,
+            /*.b_equal_seqs =*/ false,
             /*.n_tokens     =*/ (uint32_t) batch.n_tokens,
             /*.n_seq_tokens =*/ (uint32_t) 1,
             /*.n_seqs       =*/ (uint32_t) batch.n_tokens,
@@ -221,6 +223,7 @@ bool llama_batch_allocr::init(
             /*.seq_id_unq   =*/ this->seq_id_unq.data(),
             /*.seq_idx      =*/ this->seq_idx.data(),
             /*.output       =*/ batch.logits,
+            /*.data         =*/ {},
         };
 
         ubatch_print(ubatch, debug);
@@ -364,39 +367,38 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
     clear();
     split_reset();
 
-    ubatches.emplace_back();
+    auto udata = std::make_shared<llama_ubatch::data_t>();
 
-    auto & ubatch = ubatches.back();
-
-    ubatch.token     .resize(n_tokens);
-    ubatch.embd      .clear();
-    ubatch.pos       .resize(n_tokens);
-    ubatch.n_seq_id  .resize(n_tokens);
-    ubatch.seq_id    .resize(n_tokens);
-    ubatch.seq_id_unq.resize(0);
-    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
-    ubatch.output    .resize(n_tokens);
+    udata->token     .resize(n_tokens);
+    udata->embd      .clear();
+    udata->pos       .resize(n_tokens);
+    udata->n_seq_id  .resize(n_tokens);
+    udata->seq_id    .resize(n_tokens);
+    udata->seq_id_unq.resize(0);
+    udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    udata->output    .resize(n_tokens);
 
     for (uint32_t s = 0; s < n_seqs; ++s) {
-        ubatch.seq_idx[s] = s;
-        ubatch.seq_id_unq.push_back(s);
+        udata->seq_idx[s] = s;
+        udata->seq_id_unq.push_back(s);
     }
 
     llama_ubatch res {
-        /*.equal_seqs   =*/ true,
+        /*.b_equal_seqs =*/ true,
         /*.n_tokens     =*/ n_tokens,
         /*.n_seq_tokens =*/ n_seq_tokens,
         /*.n_seqs       =*/ n_seqs,
         /*.n_seqs_unq   =*/ n_seqs,
 
-        /*.token        =*/ ubatch.token.data(),
+        /*.token        =*/ udata->token.data(),
         /*.embd         =*/ nullptr,
-        /*.pos          =*/ ubatch.pos.data(),
-        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
-        /*.seq_id       =*/ ubatch.seq_id.data(),
-        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
-        /*.seq_idx      =*/ ubatch.seq_idx.data(),
-        /*.output       =*/ ubatch.output.data(),
+        /*.pos          =*/ udata->pos.data(),
+        /*.n_seq_id     =*/ udata->n_seq_id.data(),
+        /*.seq_id       =*/ udata->seq_id.data(),
+        /*.seq_id_unq   =*/ udata->seq_id_unq.data(),
+        /*.seq_idx      =*/ udata->seq_idx.data(),
+        /*.output       =*/ udata->output.data(),
+        /*.data         =*/ std::move(udata),
     };
 
     return res;
@@ -437,8 +439,6 @@ void llama_batch_allocr::split_reset() {
 
     used.clear();
     used.resize(get_n_tokens(), false);
-
-    ubatches.clear();
 }
 
 llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
@@ -653,78 +653,77 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
     assert(n_tokens%n_seqs == 0);
 
-    ubatches.emplace_back();
-
-    auto & ubatch = ubatches.back();
+    auto udata = std::make_shared<llama_ubatch::data_t>();
 
     const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;
 
     const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
     const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_cur;
 
-    ubatch.token     .resize(n_tokens);
-    ubatch.embd      .resize(n_embd_all);
-    ubatch.pos       .resize(n_pos_all);
-    ubatch.n_seq_id  .resize(n_tokens);
-    ubatch.seq_id    .resize(n_tokens);
-    ubatch.seq_id_unq.resize(0);
-    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
-    ubatch.output    .resize(n_tokens);
+    udata->token     .resize(n_tokens);
+    udata->embd      .resize(n_embd_all);
+    udata->pos       .resize(n_pos_all);
+    udata->n_seq_id  .resize(n_tokens);
+    udata->seq_id    .resize(n_tokens);
+    udata->seq_id_unq.resize(0);
+    udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    udata->output    .resize(n_tokens);
 
     seq_set_t seq_set_unq;
 
     for (size_t i = 0; i < idxs.size(); ++i) {
         if (batch.token) {
-            ubatch.token[i] = batch.token[idxs[i]];
+            udata->token[i] = batch.token[idxs[i]];
         }
 
         if (batch.embd) {
-            memcpy(ubatch.embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
+            memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
         }
 
         for (int j = 0; j < n_pos_cur; ++j) {
-            ubatch.pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
+            udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
         }
 
-        ubatch.n_seq_id[i] = batch.n_seq_id[idxs[i]];
-        ubatch.seq_id[i]   = batch.seq_id[idxs[i]];
-        ubatch.output[i]   = batch.logits[idxs[i]];
+        udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
+        udata->seq_id[i]   = batch.seq_id[idxs[i]];
+        udata->output[i]   = batch.logits[idxs[i]];
 
-        for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {
-            seq_set_unq.set(ubatch.seq_id[i][s]);
+        for (int s = 0; s < udata->n_seq_id[i]; ++s) {
+            seq_set_unq.set(udata->seq_id[i][s]);
         }
 
-        if (ubatch.output[i]) {
+        if (udata->output[i]) {
             out_ids.push_back(idxs[i]);
         }
     }
 
     for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_set_unq.test(s)) {
-            ubatch.seq_idx[s] = ubatch.seq_id_unq.size();
-            ubatch.seq_id_unq.push_back(s);
+            udata->seq_idx[s] = udata->seq_id_unq.size();
+            udata->seq_id_unq.push_back(s);
         }
     }
 
     llama_ubatch res {
-        /*.equal_seqs   =*/ equal_seqs,
+        /*.b_equal_seqs =*/ equal_seqs,
         /*.n_tokens     =*/ n_tokens,
         /*.n_seq_tokens =*/ n_tokens/n_seqs,
         /*.n_seqs       =*/ n_seqs,
-        /*.n_seqs_unq   =*/ (uint32_t) ubatch.seq_id_unq.size(),
+        /*.n_seqs_unq   =*/ (uint32_t) udata->seq_id_unq.size(),
 
-        /*.token        =*/ batch.token ? ubatch.token.data() : nullptr,
-        /*.embd         =*/ batch.embd ? ubatch.embd.data() : nullptr,
-        /*.pos          =*/ ubatch.pos.data(),
-        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
-        /*.seq_id       =*/ ubatch.seq_id.data(),
-        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
-        /*.seq_idx      =*/ ubatch.seq_idx.data(),
-        /*.output       =*/ ubatch.output.data(),
+        /*.token        =*/ batch.token ? udata->token.data() : nullptr,
+        /*.embd         =*/ batch.embd ? udata->embd.data() : nullptr,
+        /*.pos          =*/ udata->pos.data(),
+        /*.n_seq_id     =*/ udata->n_seq_id.data(),
+        /*.seq_id       =*/ udata->seq_id.data(),
+        /*.seq_id_unq   =*/ udata->seq_id_unq.data(),
+        /*.seq_idx      =*/ udata->seq_idx.data(),
+        /*.output       =*/ udata->output.data(),
+        /*.data         =*/ std::move(udata),
     };
 
     if (debug > 0) {
-        LLAMA_LOG_DEBUG("%s: added ubatch %d to split:\n", __func__, (int) ubatches.size() - 1);
+        LLAMA_LOG_DEBUG("%s: added ubatch to split:\n", __func__);
 
         ubatch_print(res, debug);
     }
@@ -734,7 +733,7 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
 
 void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) {
     if (debug > 0) {
-        LLAMA_LOG_DEBUG("%s:   equal_seqs   = %d\n", __func__, ubatch.equal_seqs);
+        LLAMA_LOG_DEBUG("%s:   equal_seqs   = %d\n", __func__, ubatch.equal_seqs());
         LLAMA_LOG_DEBUG("%s:   n_tokens     = %d\n", __func__, ubatch.n_tokens);
         LLAMA_LOG_DEBUG("%s:   n_seq_tokens = %d\n", __func__, ubatch.n_seq_tokens);
         LLAMA_LOG_DEBUG("%s:   n_seqs       = %d\n", __func__, ubatch.n_seqs);

src/llama-batch.h

@@ -8,12 +8,17 @@
 #include <vector>
 #include <set>
 #include <bitset>
+#include <memory>
 #include <unordered_map>
 
 // keep this struct lightweight
-// it points to data in `llama_batch_allocr`
 struct llama_ubatch {
-    bool equal_seqs;
+    bool equal_seqs() const {
+        return b_equal_seqs != 0;
+    }
+
+    uint32_t b_equal_seqs; // note: this is a boolean, but we use an int32_t for alignment
+                           //       otherwise address sanitizer complains
     // TODO: whole_seqs for embeddings?
 
     uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
@@ -34,6 +39,20 @@ struct llama_ubatch {
     llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id
     int32_t      *  seq_idx;    // [LLAMA_MAX_SEQ]    | -   | seq_idx
     int8_t       *  output;     // [n_tokens]         | i   | -
+
+    struct data_t {
+        std::vector<llama_token>    token;
+        std::vector<float>          embd;
+        std::vector<llama_pos>      pos;
+        std::vector<int32_t>        n_seq_id;
+        std::vector<llama_seq_id *> seq_id;
+        std::vector<llama_seq_id>   seq_id_unq;
+        std::vector<int32_t>        seq_idx;
+        std::vector<int8_t>         output;
+    };
+
+    // the llama_ubatch pointers above point to this data if set. otherwise - points to non-owning data
+    std::shared_ptr<data_t> data;
 };
 
 // a helper for sanitizing, fulfilling and splitting a batch
@@ -117,7 +136,7 @@ private:
     using seq_cpl_t = std::vector<bool>;
 
     // helper flag to quickly determine if there are any coupled sequences in the batch
-    bool has_cpl;
+    bool has_cpl = false;
 
     std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
     std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
@@ -137,20 +156,5 @@ private:
     // used[i] indicates if token i has already been used in a previous ubatch
     std::vector<bool> used;
 
-    // llama_ubatch points to this data:
-    struct ubatch {
-        std::vector<llama_token>    token;
-        std::vector<float>          embd;
-        std::vector<llama_pos>      pos;
-        std::vector<int32_t>        n_seq_id;
-        std::vector<llama_seq_id *> seq_id;
-        std::vector<llama_seq_id>   seq_id_unq;
-        std::vector<int32_t>        seq_idx;
-        std::vector<int8_t>         output;
-    };
-
-    // current splitting state:
-    std::vector<ubatch> ubatches;
-
     int debug;
 };

src/llama-chat.cpp

@@ -56,6 +56,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
+    { "exaone4",           LLM_CHAT_TEMPLATE_EXAONE_4          },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
@@ -168,6 +169,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     } else if (tmpl_contains(LU8("<｜Assistant｜>")) && tmpl_contains(LU8("<｜User｜>")) && tmpl_contains(LU8("<｜end▁of▁sentence｜>"))) {
         return LLM_CHAT_TEMPLATE_DEEPSEEK_3;
     } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) {
+        if (tmpl_contains("[|tool|]")) {
+            return LLM_CHAT_TEMPLATE_EXAONE_4;
+        }
         // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
         // EXAONE-3.0-7.8B-Instruct
         return LLM_CHAT_TEMPLATE_EXAONE_3;
@@ -532,6 +536,22 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "[|assistant|]";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_4) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "[|system|]" << trim(message->content) << "[|endofturn|]\n";
+            } else if (role == "user") {
+                ss << "[|user|]" << trim(message->content) << "\n";
+            } else if (role == "assistant") {
+                ss << "[|assistant|]" << trim(message->content) << "[|endofturn|]\n";
+            } else if (role == "tool") {
+                ss << "[|tool|]" << trim(message->content) << "[|endofturn|]\n";
+            }
+        }
+        if (add_ass) {
+            ss << "[|assistant|]";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
         // this template requires the model to have "\n\n" as EOT token
         for (size_t i = 0; i < chat.size(); i++) {

src/llama-chat.h

@@ -35,6 +35,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
+    LLM_CHAT_TEMPLATE_EXAONE_4,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,

src/llama-context.cpp

@@ -105,7 +105,7 @@ llama_context::llama_context(
 
     {
         const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-        const bool supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0;
+        const bool supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : false;
 
         if (!supports_set_rows && !cparams.kv_unified) {
             LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);
@@ -238,8 +238,8 @@ llama_context::llama_context(
 
         LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);
 
-        // buffer used to store the computation graph and the tensor meta data
-        buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
+        gf_res_prev.reset(new llm_graph_result(max_nodes));
+        gf_res_reserve.reset(new llm_graph_result(max_nodes));
 
         // TODO: move these checks to ggml_backend_sched
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
@@ -403,10 +403,6 @@ ggml_backend_sched_t llama_context::get_sched() const {
     return sched.get();
 }
 
-ggml_context * llama_context::get_ctx_compute() const {
-    return ctx_compute.get();
-}
-
 uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
@@ -478,6 +474,11 @@ bool llama_context::kv_self_update(bool optimize) {
                 }
         }
 
+        // reset the previous graph result to make sure that it won't be reused
+        // TODO: change the mctx->apply() to return information if a graph reserve is needed
+        //       reset the graph result only if the memory module did reset the scheduler
+        gf_res_prev->reset();
+
         if (!mctx->apply()) {
             LLAMA_LOG_ERROR("%s: failed to apply memory update\n", __func__);
         }
@@ -693,38 +694,59 @@ bool llama_context::apply_adapter_cvec(
     return cvec.apply(model, data, len, n_embd, il_start, il_end);
 }
 
-llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_context_i * mctx, ggml_status & ret) {
+llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_context_i * mctx, ggml_status & ret) {
     if (mctx && !mctx->apply()) {
         LLAMA_LOG_ERROR("%s: failed to apply memory context\n", __func__);
         ret = GGML_STATUS_FAILED;
         return nullptr;
     }
 
-    auto * gf = graph_init();
-    if (!gf) {
-        LLAMA_LOG_ERROR("%s: failed to initialize graph\n", __func__);
-        ret = GGML_STATUS_FAILED;
-        return nullptr;
+    auto * res = gf_res_prev.get();
+    auto * gf  = res->get_gf();
+
+    // the new graph parameters
+    // in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
+    const auto gparams = graph_params(res, ubatch, mctx, gtype);
+
+    if (res->can_reuse(gparams)) {
+        //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
+
+        n_reused++;
+    } else {
+        res->reset();
+
+        ggml_backend_sched_reset(sched.get());
+        ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
+
+        //const auto t_start_us = ggml_time_us();
+
+        gf = model.build_graph(gparams);
+
+        //LLAMA_LOG_INFO("graph build time: %.3f ms\n", (ggml_time_us() - t_start_us)/1000.0);
+
+        if (!gf) {
+            LLAMA_LOG_ERROR("%s: failed to initialize graph\n", __func__);
+            ret = GGML_STATUS_FAILED;
+            return nullptr;
+        }
+
+        if (!ggml_backend_sched_alloc_graph(sched.get(), gf)) {
+            LLAMA_LOG_ERROR("%s: failed to allocate graph\n", __func__);
+            ret = GGML_STATUS_ALLOC_FAILED;
+            return nullptr;
+        }
     }
 
-    auto res = graph_build(ctx_compute.get(), gf, ubatch, gtype, mctx);
-    if (!res) {
-        LLAMA_LOG_ERROR("%s: failed to build graph\n", __func__);
-        ret = GGML_STATUS_FAILED;
-        return nullptr;
+    // set the input data for the input tensors
+    {
+        //const auto t_start_us = ggml_time_us();
+
+        res->set_inputs(&ubatch);
+
+        //LLAMA_LOG_INFO("graph set inputs time: %.3f ms\n", (ggml_time_us() - t_start_us)/1000.0);
     }
 
-    // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
-
-    if (!ggml_backend_sched_alloc_graph(sched.get(), gf)) {
-        LLAMA_LOG_ERROR("%s: failed to allocate graph\n", __func__);
-        ret = GGML_STATUS_ALLOC_FAILED;
-        return nullptr;
-    }
-
-    res->set_inputs(&ubatch);
-
-    const auto status = graph_compute(gf, ubatch.n_tokens > 1);
+    const auto status = graph_compute(res->get_gf(), ubatch.n_tokens > 1);
     if (status != GGML_STATUS_SUCCESS) {
         LLAMA_LOG_ERROR("%s: failed to compute graph, compute status: %d\n", __func__, status);
         ret = status;
@@ -785,9 +807,6 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     n_outputs = n_tokens;
 
-    ggml_backend_sched_reset(sched.get());
-    ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
-
     const auto causal_attn_org = cparams.causal_attn;
 
     // always use non-causal attention for encoder graphs
@@ -796,7 +815,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     cparams.causal_attn = false;
 
     ggml_status status;
-    const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_ENCODER, nullptr, status);
+    const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_ENCODER, nullptr, status);
 
     cparams.causal_attn = causal_attn_org;
 
@@ -872,10 +891,6 @@ int llama_context::encode(const llama_batch & batch_inp) {
         }
     }
 
-    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-    // overlap with device computation.
-    ggml_backend_sched_reset(sched.get());
-
     // TODO: hacky solution
     if (model.arch == LLM_ARCH_T5 && t_embd) {
         //cross.t_embd = t_embd;
@@ -1033,11 +1048,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
             n_outputs = n_outputs_new;
         }
 
-        ggml_backend_sched_reset(sched.get());
-        ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
-
         ggml_status status;
-        const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
+        const auto * res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
 
         if (!res) {
             // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
@@ -1218,10 +1230,6 @@ int llama_context::decode(const llama_batch & batch_inp) {
     // wait for the computation to finish (automatically done when obtaining the model output)
     //synchronize();
 
-    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-    // overlap with device computation.
-    ggml_backend_sched_reset(sched.get());
-
     return 0;
 }
 
@@ -1303,20 +1311,12 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
 // graph
 //
 
-int32_t llama_context::graph_max_nodes() const {
-    return std::max<int32_t>(65536, 5*model.n_tensors());
+uint32_t llama_context::graph_max_nodes() const {
+    return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }
 
-ggml_cgraph * llama_context::graph_init() {
-    ggml_init_params params = {
-        /*.mem_size   =*/ buf_compute_meta.size(),
-        /*.mem_buffer =*/ buf_compute_meta.data(),
-        /*.no_alloc   =*/ true,
-    };
-
-    ctx_compute.reset(ggml_init(params));
-
-    return ggml_new_graph_custom(ctx_compute.get(), graph_max_nodes(), false);
+llm_graph_result * llama_context::get_gf_res_reserve() const {
+    return static_cast<llm_graph_result *>(gf_res_reserve.get());
 }
 
 ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) {
@@ -1329,6 +1329,11 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
         LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs);
     }
 
+    ggml_backend_sched_reset(sched.get());
+
+    // when the scheduler is reset, we cannnot reuse the old graph, so we reset the previous graph result to prevent that
+    gf_res_prev->reset();
+
     // store the n_outputs as it is, and restore it afterwards
     // TODO: not sure if needed, might simplify in the future by removing this
     const auto save_n_outputs = this->n_outputs;
@@ -1338,18 +1343,16 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     llama_batch_allocr balloc(model.hparams.n_pos_per_embd());
     llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
 
-    auto * gf = graph_init();
-    auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mctx);
+    auto * res = gf_res_reserve.get();
+
+    const auto gparams = graph_params(res, ubatch, mctx, LLM_GRAPH_TYPE_DEFAULT);
+
+    res->reset();
+
+    auto * gf = model.build_graph(gparams);
 
     this->n_outputs = save_n_outputs;
 
-    if (!res) {
-        LLAMA_LOG_ERROR("%s: failed to build worst-case graph\n", __func__);
-        return nullptr;
-    }
-
-    ggml_backend_sched_reset(sched.get());
-
     // initialize scheduler with the specified graph
     if (!ggml_backend_sched_reserve(sched.get(), gf)) {
         LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
@@ -1359,28 +1362,27 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
     return gf;
 }
 
-llm_graph_result_ptr llama_context::graph_build(
-                      ggml_context * ctx,
-                       ggml_cgraph * gf,
-                const llama_ubatch & ubatch,
-                    llm_graph_type   gtype,
-      const llama_memory_context_i * mctx) {
-    return model.build_graph(
-            {
-                /*.ctx         =*/ ctx,
-                /*.arch        =*/ model.arch,
-                /*.hparams     =*/ model.hparams,
-                /*.cparams     =*/ cparams,
-                /*.ubatch      =*/ ubatch,
-                /*.sched       =*/ sched.get(),
-                /*.backend_cpu =*/ backend_cpu,
-                /*.cvec        =*/ &cvec,
-                /*.loras       =*/ &loras,
-                /*.mctx        =*/ mctx,
-                /*.cross       =*/ &cross,
-                /*.n_outputs   =*/ n_outputs,
-                /*.cb          =*/ graph_get_cb(),
-            }, gf, gtype);
+llm_graph_params llama_context::graph_params(
+                        llm_graph_result * res,
+                      const llama_ubatch & ubatch,
+            const llama_memory_context_i * mctx,
+            llm_graph_type   gtype) const {
+    return {
+        /*.arch        =*/ model.arch,
+        /*.hparams     =*/ model.hparams,
+        /*.cparams     =*/ cparams,
+        /*.ubatch      =*/ ubatch,
+        /*.gtype       =*/ gtype,
+        /*.sched       =*/ sched.get(),
+        /*.backend_cpu =*/ backend_cpu,
+        /*.cvec        =*/ &cvec,
+        /*.loras       =*/ &loras,
+        /*.mctx        =*/ mctx,
+        /*.cross       =*/ &cross,
+        /*.n_outputs   =*/ n_outputs,
+        /*.cb          =*/ graph_get_cb(),
+        /*.res         =*/ res,
+    };
 }
 
 ggml_status llama_context::graph_compute(
@@ -1958,6 +1960,7 @@ llama_perf_context_data llama_context::perf_get_data() const {
     data.t_eval_ms   = 1e-3 * t_eval_us;
     data.n_p_eval    = std::max(1, n_p_eval);
     data.n_eval      = std::max(1, n_eval);
+    data.n_reused    = std::max(0, n_reused);
 
     return data;
 }
@@ -1966,6 +1969,7 @@ void llama_context::perf_reset() {
     t_start_us  = ggml_time_us();
     t_eval_us   = n_eval = 0;
     t_p_eval_us = n_p_eval = 0;
+    n_reused    = 0;
 }
 
 //
@@ -2092,8 +2096,13 @@ void llama_context::opt_epoch_iter(
                 break;
             }
 
-            auto * gf = graph_init();
-            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mctx.get());
+            auto * res = gf_res_prev.get();
+
+            const auto gparams = graph_params(res, ubatch, mctx.get(), LLM_GRAPH_TYPE_DEFAULT);
+
+            res->reset();
+
+            auto * gf = model.build_graph(gparams);
 
             struct ggml_context * ctx_compute_opt;
             {
@@ -2836,6 +2845,7 @@ void llama_perf_context_print(const llama_context * ctx) {
     LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
     LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
+    LLAMA_LOG_INFO("%s:    graphs reused = %10d\n", __func__, data.n_reused);
 }
 
 void llama_perf_context_reset(llama_context * ctx) {

src/llama-context.h

@@ -35,8 +35,6 @@ struct llama_context {
 
     ggml_backend_sched_t get_sched() const;
 
-    ggml_context * get_ctx_compute() const;
-
     uint32_t n_ctx()         const;
     uint32_t n_ctx_per_seq() const;
     uint32_t n_batch()       const;
@@ -96,7 +94,7 @@ struct llama_context {
     // if memory_context is provided, it will be applied first to the context's memory
     // ret contains the status of the graph computation
     // returns nullptr only if ret != GGML_STATUS_SUCCESS
-    llm_graph_result_ptr process_ubatch(
+    llm_graph_result * process_ubatch(
                 const llama_ubatch & ubatch,
                     llm_graph_type   gtype,
             llama_memory_context_i * mctx,
@@ -188,10 +186,10 @@ private:
     //
 
 public:
-    int32_t graph_max_nodes() const;
+    uint32_t graph_max_nodes() const;
 
-    // zero-out inputs and create the ctx_compute for the compute graph
-    ggml_cgraph * graph_init();
+    // can reuse the llm_graph_result instance of the context (for example to update a memory module)
+    llm_graph_result * get_gf_res_reserve() const;
 
     // returns the result of ggml_backend_sched_graph_compute_async execution
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
@@ -200,12 +198,11 @@ public:
     ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
 
 private:
-    llm_graph_result_ptr graph_build(
-                      ggml_context * ctx,
-                       ggml_cgraph * gf,
-                const llama_ubatch & ubatch,
-                    llm_graph_type   gtype,
-      const llama_memory_context_i * mctx);
+    llm_graph_params graph_params(
+                        llm_graph_result * res,
+                      const llama_ubatch & ubatch,
+            const llama_memory_context_i * mctx,
+                          llm_graph_type   gtype) const;
 
     llm_graph_cb graph_get_cb() const;
 
@@ -258,8 +255,6 @@ private:
     ggml_backend_t backend_cpu = nullptr;
     std::vector<ggml_backend_ptr> backends;
 
-    ggml_context_ptr ctx_compute;
-
     // training
     ggml_opt_context_t opt_ctx = nullptr;
 
@@ -275,8 +270,8 @@ private:
     std::vector<ggml_backend_t>             backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
 
-    // memory buffers used to evaluate the model
-    std::vector<uint8_t> buf_compute_meta;
+    llm_graph_result_ptr gf_res_prev;
+    llm_graph_result_ptr gf_res_reserve;
 
     // host buffer for the model output (logits and embeddings)
     ggml_backend_buffer_ptr buf_output;
@@ -294,4 +289,6 @@ private:
 
     mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
     mutable int32_t n_eval   = 0; // number of eval calls
+
+    mutable int32_t n_reused = 0; // number of times the previous graph was reused
 };

src/llama-graph.cpp

@@ -28,6 +28,15 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
+    bool res = true;
+
+    res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
+    res &= (!embd   && !params.ubatch.embd)  || (embd   &&   embd->ne[0] == params.ubatch.n_tokens);
+
+    return res;
+}
+
 void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && pos) {
         const int64_t n_tokens = ubatch->n_tokens;
@@ -50,6 +59,14 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_pos::can_reuse(const llm_graph_params & params) {
+    bool res = true;
+
+    res &= pos->ne[0] == params.ubatch.n_tokens;
+
+    return res;
+}
+
 void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && attn_scale) {
         const int64_t n_tokens = ubatch->n_tokens;
@@ -71,7 +88,7 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
         const int64_t n_tokens = ubatch->n_tokens;
 
         GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer));
-        GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+        GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
 
         int32_t * data = (int32_t *) pos_bucket->data;
 
@@ -118,6 +135,14 @@ void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+bool llm_graph_input_out_ids::can_reuse(const llm_graph_params & params) {
+    bool res = true;
+
+    res &= n_outputs == params.n_outputs;
+
+    return res;
+}
+
 void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
     if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
         const int64_t n_tokens     = ubatch->n_tokens;
@@ -287,6 +312,24 @@ void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
     mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
 }
 
+bool llm_graph_input_attn_kv_unified::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_unified_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
+  //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
+    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+
+    res &= mctx->get_supports_set_rows(); // TODO: tmp
+
+    return res;
+}
+
 void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
     mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
     mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
@@ -299,6 +342,30 @@ void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch
     mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
 }
 
+bool llm_graph_input_attn_kv_unified_iswa::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_unified_iswa_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
+  //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
+  //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+    res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
+    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+
+    res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
+    res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+
+    res &= mctx->get_base()->get_supports_set_rows(); // TODO: tmp
+
+    return res;
+}
+
 void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
     GGML_ASSERT(cross_kq_mask);
 
@@ -306,7 +373,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
     const int64_t n_tokens = ubatch->n_tokens;
 
     GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
-    GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+    GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
 
     float * data = (float *) cross_kq_mask->data;
 
@@ -340,6 +407,91 @@ void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
     inp_rs->set_input(ubatch);
 }
 
+//
+// llm_graph_result
+//
+
+llm_graph_result::llm_graph_result(int64_t max_nodes) : max_nodes(max_nodes) {
+    reset();
+
+    const char * LLAMA_GRAPH_RESULT_DEBUG = getenv("LLAMA_GRAPH_RESULT_DEBUG");
+    debug = LLAMA_GRAPH_RESULT_DEBUG ? atoi(LLAMA_GRAPH_RESULT_DEBUG) : 0;
+}
+
+int64_t llm_graph_result::get_max_nodes() const {
+    return max_nodes;
+}
+
+void llm_graph_result::reset() {
+    t_tokens      = nullptr;
+    t_logits      = nullptr;
+    t_embd        = nullptr;
+    t_embd_pooled = nullptr;
+
+    params = {};
+
+    inputs.clear();
+
+    buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
+
+    ggml_init_params params = {
+        /*.mem_size   =*/ buf_compute_meta.size(),
+        /*.mem_buffer =*/ buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
+    };
+
+    ctx_compute.reset(ggml_init(params));
+
+    gf = ggml_new_graph_custom(ctx_compute.get(), max_nodes, false);
+}
+
+void llm_graph_result::set_inputs(const llama_ubatch * ubatch) {
+    for (auto & input : inputs) {
+        input->set_input(ubatch);
+    }
+}
+
+bool llm_graph_result::can_reuse(const llm_graph_params & params) {
+    if (!this->params.allow_reuse(params)) {
+        if (debug > 1) {
+            LLAMA_LOG_DEBUG("%s: cannot reuse graph due to incompatible graph parameters\n", __func__);
+        }
+
+        return false;
+    }
+
+    if (debug > 1) {
+        LLAMA_LOG_DEBUG("%s: checking compatibility of %d inputs:\n", __func__, (int) inputs.size());
+    }
+
+    bool res = true;
+
+    for (auto & input : inputs) {
+        const bool cur = input->can_reuse(params);
+
+        if (debug > 1) {
+            LLAMA_LOG_DEBUG("%s: can_reuse = %d\n", "placeholder", cur);
+        }
+
+        res = res && cur;
+    }
+
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: can reuse graph = %d\n", __func__, res);
+    }
+
+    return res;
+}
+
+llm_graph_input_i * llm_graph_result::add_input(llm_graph_input_ptr input) {
+    inputs.emplace_back(std::move(input));
+    return inputs.back().get();
+}
+
+void llm_graph_result::set_params(const llm_graph_params & params) {
+    this->params = params;
+}
+
 //
 // llm_graph_context
 //
@@ -374,7 +526,6 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     n_ctx_orig       (cparams.n_ctx_orig_yarn),
     pooling_type     (cparams.pooling_type),
     rope_type        (hparams.rope_type),
-    ctx0             (params.ctx),
     sched            (params.sched),
     backend_cpu      (params.backend_cpu),
     cvec             (params.cvec),
@@ -382,7 +533,10 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     mctx             (params.mctx),
     cross            (params.cross),
     cb_func          (params.cb),
-    res              (std::make_unique<llm_graph_result>()) {
+    res              (params.res),
+    ctx0             (res->get_ctx()),
+    gf               (res->get_gf()) {
+        res->set_params(params);
     }
 
 void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
@@ -753,20 +907,28 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(cur, "ffn_moe_weighted", il);
     }
 
-    // aggregate experts
-    ggml_tensor * moe_out = nullptr;
-    for (int i = 0; i < n_expert_used; ++i) {
-        ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens,
-                experts->nb[2], i*experts->nb[1]);
+    ggml_tensor * cur_experts[LLAMA_MAX_EXPERTS] = { nullptr };
 
-        if (i == 0) {
-            moe_out = cur_expert;
-        } else {
-            moe_out = ggml_add(ctx0, moe_out, cur_expert);
-        }
+    assert(n_expert_used > 0);
+
+    // order the views before the adds
+    for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
+        cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]);
+
+        ggml_build_forward_expand(gf, cur_experts[i]);
     }
 
-    if (n_expert_used == 1) {
+    // aggregate experts
+    // note: here we explicitly use hparams.n_expert_used instead of n_expert_used
+    //       to avoid potentially a large number of add nodes during warmup
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/14753
+    ggml_tensor * moe_out = cur_experts[0];
+
+    for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
+        moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
+    }
+
+    if (hparams.n_expert_used == 1) {
         // avoid returning a non-contiguous tensor
         moe_out = ggml_cont(ctx0, moe_out);
     }
@@ -972,7 +1134,6 @@ ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_t
 }
 
 ggml_tensor * llm_graph_context::build_attn_mha(
-         ggml_cgraph * gf,
          ggml_tensor * q,
          ggml_tensor * k,
          ggml_tensor * v,
@@ -1106,7 +1267,6 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_no_cache * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1127,14 +1287,14 @@ ggml_tensor * llm_graph_context::build_attn(
     const auto & kq_mask = inp->get_kq_mask();
 
     // [TAG_NO_CACHE_PAD]
-    // TODO: if ubatch.equal_seqs == true, we can split the three tensors below into ubatch.n_seqs_unq streams
-    assert(ubatch.equal_seqs == false);
+    // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
+    assert(!ubatch.equal_seqs());
 
     ggml_tensor * q = q_cur;
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1190,7 +1350,6 @@ llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified()
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_unified * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1223,7 +1382,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1243,7 +1402,6 @@ ggml_tensor * llm_graph_context::build_attn(
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_unified_iswa * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1290,7 +1448,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1323,7 +1481,6 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_cross * inp,
-        ggml_cgraph * gf,
         ggml_tensor * wo,
         ggml_tensor * wo_b,
         ggml_tensor * q_cur,
@@ -1345,7 +1502,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
 
     if (wo) {
@@ -1403,7 +1560,6 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
 }
 
 ggml_tensor * llm_graph_context::build_rs(
-        ggml_cgraph * gf,
         ggml_tensor * s,
         ggml_tensor * state_copy,
             int32_t   state_size,
@@ -1461,21 +1617,19 @@ llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
 
 ggml_tensor * llm_graph_context::build_rs(
         llm_graph_input_rs * inp,
-        ggml_cgraph * gf,
         ggml_tensor * s,
             int32_t   state_size,
             int32_t   n_seqs,
         const llm_graph_get_rows_fn & get_state_rows) const {
     const auto * kv_state = inp->mctx;
 
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
+    return build_rs(s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
 }
 
 ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
     llm_graph_input_rs * inp,
-           ggml_cgraph * gf,
     const llama_ubatch & ubatch,
-                 int   il) const {
+                   int   il) const {
     const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
     const auto token_shift_count = hparams.token_shift_count;
@@ -1485,7 +1639,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
     ggml_tensor * token_shift_all = mctx_cur->get_r_l(il);
 
     ggml_tensor * token_shift = build_rs(
-            inp, gf, token_shift_all,
+            inp, token_shift_all,
             hparams.n_embd_r(), n_seqs);
 
     token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);
@@ -1525,7 +1679,6 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
 }
 
 void llm_graph_context::build_pooling(
-        ggml_cgraph * gf,
         ggml_tensor * cls,
         ggml_tensor * cls_b,
         ggml_tensor * cls_out,

src/llama-graph.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "llama-arch.h"
+#include "llama-batch.h"
 #include "llama-hparams.h"
 #include "llama-adapter.h"
 
@@ -14,7 +15,6 @@ struct ggml_cgraph;
 struct ggml_context;
 struct ggml_tensor;
 
-struct llama_ubatch;
 struct llama_cparams;
 
 struct llama_memory_context_i;
@@ -69,6 +69,8 @@ struct llama_cross {
     std::vector<std::set<llama_seq_id>> seq_ids_enc;
 };
 
+struct llm_graph_params;
+
 //
 // llm_graph_input
 //
@@ -78,11 +80,19 @@ public:
     virtual ~llm_graph_input_i() = default;
 
     virtual void set_input(const llama_ubatch * ubatch) = 0;
+
+    // return true if the resulting input tensors using the provided graph parameters would be
+    //   the same as the previous input tensors that we have currently stored in the object
+    virtual bool can_reuse(const llm_graph_params & params) {
+        // returning false here by default will prevent from reusing the graph if the check
+        //   for the input type has not been implemented yet
+        GGML_UNUSED(params);
+        return false;
+    }
 };
 
 using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
 
-
 class llm_graph_input_embd : public llm_graph_input_i {
 public:
     llm_graph_input_embd()          = default;
@@ -90,6 +100,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
 };
@@ -101,6 +113,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * pos = nullptr; // I32 [n_batch]
 
     const uint32_t n_pos_per_embd = 1;
@@ -154,17 +168,19 @@ public:
     llm_graph_input_out_ids(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            int32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
+            uint32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
     virtual ~llm_graph_input_out_ids() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * out_ids; // I32 [n_outputs]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const int32_t n_outputs;
+    const uint32_t n_outputs;
 };
 
 class llm_graph_input_mean : public llm_graph_input_i {
@@ -249,6 +265,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * get_k_idxs() const { return self_k_idxs; }
     ggml_tensor * get_v_idxs() const { return self_v_idxs; }
 
@@ -280,6 +298,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * get_k_idxs()     const { return self_k_idxs; }
     ggml_tensor * get_v_idxs()     const { return self_v_idxs; }
     ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; }
@@ -351,65 +371,20 @@ public:
 // along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc.
 //   these are used by the llama_context to extact the relevant data, based on the compute parameters
 
-class llm_graph_result_i {
-public:
-    virtual ~llm_graph_result_i() = default;
-
-    virtual ggml_tensor * get_tokens()      = 0;
-    virtual ggml_tensor * get_logits()      = 0;
-    virtual ggml_tensor * get_embd()        = 0;
-    virtual ggml_tensor * get_embd_pooled() = 0;
-
-    virtual void set_inputs(const llama_ubatch * ubatch) = 0;
-};
-
-using llm_graph_result_ptr = std::unique_ptr<llm_graph_result_i>;
-
-
-class llm_graph_result : public llm_graph_result_i {
-public:
-    virtual ~llm_graph_result() = default;
-
-    ggml_tensor * get_tokens()      override { return t_tokens; }
-    ggml_tensor * get_logits()      override { return t_logits; }
-    ggml_tensor * get_embd()        override { return t_embd; }
-    ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
-
-    void set_inputs(const llama_ubatch * ubatch) override {
-        for (auto & input : inputs) {
-            input->set_input(ubatch);
-        }
-    }
-
-    llm_graph_input_i * add_input(llm_graph_input_ptr input) {
-        inputs.emplace_back(std::move(input));
-        return inputs.back().get();
-    }
-
-    // important graph nodes
-    ggml_tensor * t_tokens      = nullptr;
-    ggml_tensor * t_logits      = nullptr;
-    ggml_tensor * t_embd        = nullptr;
-    ggml_tensor * t_embd_pooled = nullptr;
-
-    std::vector<llm_graph_input_ptr> inputs;
-};
-
-//
-// llm_graph_context
-//
-
 // callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
 using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;
 
+class llm_graph_result;
+
 struct llm_graph_params {
-    ggml_context * ctx;
+    llm_arch arch = LLM_ARCH_UNKNOWN;
 
-    const llm_arch arch;
+    llama_hparams hparams;
+    llama_cparams cparams;
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
-    const llama_ubatch  & ubatch;
+    llama_ubatch ubatch; // note: intentionally make a copy
+
+    llm_graph_type gtype;
 
     ggml_backend_sched_t sched;
     ggml_backend_t backend_cpu;
@@ -421,9 +396,117 @@ struct llm_graph_params {
 
     uint32_t n_outputs;
 
-    const llm_graph_cb & cb;
+    llm_graph_cb cb;
+
+    llm_graph_result * res;
+
+    // return true if the "other" params would result in a graph with the same topology as with the current params
+    //   having the same topology allows us to reuse the graph in some cases
+    bool allow_reuse(const llm_graph_params & other) const {
+        // first check the ubatch
+        bool can_reuse_ubatch =
+            ubatch.equal_seqs() == other.ubatch.equal_seqs() &&
+            ubatch.n_tokens     == other.ubatch.n_tokens &&
+            ubatch.n_seq_tokens == other.ubatch.n_seq_tokens &&
+            ubatch.n_seqs       == other.ubatch.n_seqs &&
+            ubatch.n_seqs_unq   == other.ubatch.n_seqs_unq &&
+            (
+                (!ubatch.token && !other.ubatch.token) ||
+                (!ubatch.embd  && !other.ubatch.embd)
+            );
+
+        if (can_reuse_ubatch && !ubatch.equal_seqs()) {
+            if (!ubatch.data) {
+                // if the old ubatch does not own it's data, then we cannot guarantee that it is still alive, and
+                //   therefore we cannot perform the sequence id check. normally should never happen
+                can_reuse_ubatch = false;
+            } else {
+                for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                    can_reuse_ubatch &= ubatch.seq_id_unq[s] == other.ubatch.seq_id_unq[s];
+                }
+            }
+        }
+
+        if (!can_reuse_ubatch) {
+            return false;
+        }
+
+        return
+            cparams.embeddings  == other.cparams.embeddings  &&
+            cparams.causal_attn == other.cparams.causal_attn &&
+            arch      == other.arch  &&
+            gtype     == other.gtype &&
+            cvec      == other.cvec  &&
+            loras     == other.loras &&
+            cross     == other.cross &&
+            n_outputs == other.n_outputs;
+    }
 };
 
+class llm_graph_result {
+public:
+    llm_graph_result(int64_t max_nodes);
+
+    virtual ~llm_graph_result() = default;
+
+    ggml_tensor * get_tokens()      const { return t_tokens; }
+    ggml_tensor * get_logits()      const { return t_logits; }
+    ggml_tensor * get_embd()        const { return t_embd; }
+    ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
+
+    ggml_cgraph  * get_gf()  const { return gf; }
+    ggml_context * get_ctx() const { return ctx_compute.get(); }
+
+    int64_t get_max_nodes() const;
+
+    void reset();
+
+    void set_inputs(const llama_ubatch * ubatch);
+
+    // try to update the existing graph result using the new graph parameters in order to reuse it
+    // this can only be done if we determine that the resulting graph using the new graph parameters
+    //   would be identical to the existing graph. in that case, we simply have to update the memory
+    //   contexts of the input tensors of the graph and we can reuse it for another computation
+    // return true if the graph was updated and can be reused
+    bool can_reuse(const llm_graph_params & params);
+
+    llm_graph_input_i * add_input(llm_graph_input_ptr input);
+
+    void set_params(const llm_graph_params & params);
+
+    // important graph nodes
+    ggml_tensor * t_tokens      = nullptr;
+    ggml_tensor * t_logits      = nullptr;
+    ggml_tensor * t_embd        = nullptr;
+    ggml_tensor * t_embd_pooled = nullptr;
+
+    std::vector<llm_graph_input_ptr> inputs;
+
+    ggml_context_ptr ctx_compute;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+
+    ggml_cgraph * gf;
+
+    int64_t max_nodes;
+
+private:
+    // keep a copy of the previous graph parameters
+    // we will use this to determine whether the graph can be reused by comparing them with the new parameters
+    // note: these are updated after constructing the new graph
+    llm_graph_params params;
+
+    // env: LLAMA_GRAPH_RESULT_DEBUG
+    int debug = 0;
+};
+
+using llm_graph_result_ptr = std::unique_ptr<llm_graph_result>;
+
+//
+// llm_graph_context
+//
+
 // used in build_rs to properly order writes and avoid unnecessary copies
 using llm_graph_get_rows_fn = std::function<ggml_tensor * (ggml_context *, ggml_tensor * states, ggml_tensor * ids)>;
 
@@ -463,8 +546,6 @@ struct llm_graph_context {
     const enum llama_pooling_type pooling_type;
     const enum llama_rope_type    rope_type;
 
-    ggml_context * ctx0 = nullptr;
-
     ggml_backend_sched_t sched;
 
     ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
@@ -476,7 +557,10 @@ struct llm_graph_context {
 
     const llm_graph_cb & cb_func;
 
-    std::unique_ptr<llm_graph_result> res;
+    llm_graph_result * res;
+
+    ggml_context * ctx0 = nullptr;
+    ggml_cgraph  * gf   = nullptr;
 
     llm_graph_context(const llm_graph_params & params);
     virtual ~llm_graph_context() = default;
@@ -562,7 +646,6 @@ struct llm_graph_context {
     //
 
     ggml_tensor * build_attn_mha(
-             ggml_cgraph * gf,
              ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
              ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
              ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
@@ -575,7 +658,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_no_cache * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -590,7 +672,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -606,7 +687,6 @@ struct llm_graph_context {
     // note: if k_cur or v_cur are not provided, they will not be stored in the memory
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified_iswa * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -621,7 +701,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_cross * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -643,7 +722,6 @@ struct llm_graph_context {
     //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
     //         `llama_memory_recurrent`
     ggml_tensor * build_rs(
-            ggml_cgraph * gf,
             ggml_tensor * s,
             ggml_tensor * state_copy,
                 int32_t   state_size,
@@ -658,7 +736,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_rs(
             llm_graph_input_rs * inp,
-            ggml_cgraph * gf,
             ggml_tensor * s,
                 int32_t   state_size,
                 int32_t   n_seqs,
@@ -666,9 +743,8 @@ struct llm_graph_context {
 
     ggml_tensor * build_rwkv_token_shift_load(
         llm_graph_input_rs * inp,
-               ggml_cgraph * gf,
         const llama_ubatch & ubatch,
-                     int   il) const;
+                       int   il) const;
 
     ggml_tensor * build_rwkv_token_shift_store(
              ggml_tensor * token_shift,
@@ -685,7 +761,6 @@ struct llm_graph_context {
     //
 
     void build_pooling(
-            ggml_cgraph * gf,
             ggml_tensor * cls,
             ggml_tensor * cls_b,
             ggml_tensor * cls_out,

src/llama-kv-cache-unified.cpp

@@ -193,7 +193,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
 
     const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0;
+    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) != 0 : 0;
 
     if (!supports_set_rows) {
         // ref: https://github.com/ggml-org/llama.cpp/pull/14363
@@ -656,14 +656,11 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
         if (hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
             ggml_backend_sched_reset(sched);
 
-            auto * gf = lctx->graph_init();
+            auto * res = lctx->get_gf_res_reserve();
 
-            auto res = build_graph_shift(lctx->get_cparams(), lctx->get_ctx_compute(), gf);
-            if (!res) {
-                LLAMA_LOG_ERROR("%s: failed to build graph for K-shift\n", __func__);
-                return updated;
-            }
+            res->reset();
 
+            auto * gf = build_graph_shift(res, lctx);
             if (!ggml_backend_sched_alloc_graph(sched, gf)) {
                 LLAMA_LOG_ERROR("%s: failed to allocate compute graph for K-shift\n", __func__);
                 return updated;
@@ -713,14 +710,11 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
 
         ggml_backend_sched_reset(sched);
 
-        auto * gf = lctx->graph_init();
+        auto * res = lctx->get_gf_res_reserve();
 
-        auto res = build_graph_defrag(lctx->get_cparams(), lctx->get_ctx_compute(), gf, dinfo);
-        if (!res) {
-            LLAMA_LOG_ERROR("%s: failed to build graph for defrag\n", __func__);
-            return updated;
-        }
+        res->reset();
 
+        auto * gf = build_graph_defrag(res, lctx, dinfo);
         if (!ggml_backend_sched_alloc_graph(sched, gf)) {
             LLAMA_LOG_ERROR("%s: failed to allocate compute graph for defrag\n", __func__);
             return updated;
@@ -1035,6 +1029,10 @@ uint32_t llama_kv_cache_unified::get_n_kv() const {
     return result;
 }
 
+bool llama_kv_cache_unified::get_supports_set_rows() const {
+    return supports_set_rows;
+}
+
 ggml_tensor * llama_kv_cache_unified::get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
@@ -1263,7 +1261,7 @@ void llama_kv_cache_unified::set_input_k_shift(ggml_tensor * dst) const {
         const auto & cells = v_cells[s];
 
         for (uint32_t i = 0; i < cells.size(); ++i) {
-            data[i] = cells.is_empty(i) ? 0 : cells.get_shift(i);
+            data[s*cells.size() + i] = cells.is_empty(i) ? 0 : cells.get_shift(i);
         }
     }
 }
@@ -1283,6 +1281,8 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
     const int64_t n_tps     = n_tokens/n_stream;
     const int64_t n_tps_pad = GGML_PAD(n_tps, GGML_KQ_MASK_PAD);
 
+    std::fill(data, data + ggml_nelements(dst), -INFINITY);
+
     // Use only the previous KV cells of the correct sequence for each token of the ubatch.
     // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
     // Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch:
@@ -1295,6 +1295,7 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
     //      xxxxx-----
     //      xxxxx-----
     // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
+    // TODO: optimize this section
     for (uint32_t h = 0; h < 1; ++h) {
         for (uint32_t s = 0; s < n_stream; ++s) {
             for (uint32_t ii = 0; ii < n_tps; ++ii) {
@@ -1306,44 +1307,31 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
 
                 const llama_pos p1 = ubatch->pos[i];
 
+                const uint64_t idst = n_kv*(h*n_stream*n_tps_pad + s*n_tps_pad + ii);
+
                 for (uint32_t j = 0; j < n_kv; ++j) {
-                    float f = 0.0f;
-
-                    bool masked = false;
-
                     if (cells.is_empty(j)) {
-                        masked = true;
-                    } else {
-                        const llama_pos p0 = cells.pos_get(j);
-
-                        // mask the token if not the same sequence
-                        masked = masked || (!cells.seq_has(j, seq_id));
-
-                        // mask future tokens
-                        masked = masked || (causal_attn && p0 > p1);
-
-                        // apply SWA if any
-                        masked = masked || (is_masked_swa(p0, p1));
-
-                        if (!masked && hparams.use_alibi) {
-                            f = -std::abs(p0 - p1);
-                        }
+                        continue;
                     }
 
-                    if (masked) {
-                        f = -INFINITY;
+                    // mask the token if not the same sequence
+                    if (!cells.seq_has(j, seq_id)) {
+                        continue;
                     }
 
-                    data[h*n_stream*n_tps_pad*n_kv + s*n_tps_pad*n_kv + ii*n_kv + j] = f;
-                }
+                    const llama_pos p0 = cells.pos_get(j);
 
-                // mask padded tokens
-                if (data) {
-                    for (uint32_t ii = n_tps; ii < n_tps_pad; ++ii) {
-                        for (uint32_t j = 0; j < n_kv; ++j) {
-                            data[h*n_stream*n_tps_pad*n_kv + s*n_tps_pad*n_kv + ii*n_kv + j] = -INFINITY;
-                        }
+                    // mask future tokens
+                    if (causal_attn && p0 > p1) {
+                        continue;
                     }
+
+                    // apply SWA if any
+                    if (is_masked_swa(p0, p1)) {
+                        continue;
+                    }
+
+                    data[idst + j] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
                 }
             }
         }
@@ -1357,7 +1345,7 @@ void llama_kv_cache_unified::set_input_pos_bucket(ggml_tensor * dst, const llama
     const auto & cells = v_cells[0];
 
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
-    GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+    GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
 
     int32_t * data = (int32_t *) dst->data;
 
@@ -1475,11 +1463,9 @@ void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) {
     }
 }
 
-llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
-        const llama_cparams & cparams,
-               ggml_context * ctx,
-                ggml_cgraph * gf) const {
-    auto res = std::make_unique<llm_graph_result>();
+ggml_cgraph * llama_kv_cache_unified::build_graph_shift(llm_graph_result * res, llama_context * lctx) const {
+    auto * ctx = res->get_ctx();
+    auto * gf  = res->get_gf();
 
     const auto & n_embd_head_k = hparams.n_embd_head_k;
   //const auto & n_embd_head_v = hparams.n_embd_head_v;
@@ -1489,6 +1475,8 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
     inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, (int64_t) get_size()*n_stream);
     ggml_set_input(inp->k_shift);
 
+    const auto & cparams = lctx->get_cparams();
+
     for (const auto & layer : layers) {
         const uint32_t il = layer.il;
 
@@ -1514,15 +1502,15 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
 
     res->add_input(std::move(inp));
 
-    return res;
+    return gf;
 }
 
-llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
-                const llama_cparams & cparams,
-                       ggml_context * ctx,
-                        ggml_cgraph * gf,
-                  const defrag_info & dinfo) const {
-    auto res = std::make_unique<llm_graph_result>();
+ggml_cgraph * llama_kv_cache_unified::build_graph_defrag(
+         llm_graph_result * res,
+            llama_context * lctx,
+        const defrag_info & dinfo) const {
+    auto * ctx = res->get_ctx();
+    auto * gf  = res->get_gf();
 
     GGML_ASSERT(n_stream == 1 && "n_stream > 1 does not support defrag");
 
@@ -1530,6 +1518,8 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
 
     const auto & ids = dinfo.ids;
 
+    const auto & cparams = lctx->get_cparams();
+
 #if 0
     // CPU defrag
     //
@@ -1666,7 +1656,7 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
     //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
 #endif
 
-    return res;
+    return gf;
 }
 
 llama_kv_cache_unified::defrag_info llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) const {
@@ -2342,6 +2332,10 @@ uint32_t llama_kv_cache_unified_context::get_n_kv() const {
     return n_kv;
 }
 
+bool llama_kv_cache_unified_context::get_supports_set_rows() const {
+    return kv->get_supports_set_rows();
+}
+
 ggml_tensor * llama_kv_cache_unified_context::get_k(ggml_context * ctx, int32_t il) const {
     return kv->get_k(ctx, il, n_kv, sinfos[i_cur]);
 }

src/llama-kv-cache-unified.h

@@ -154,6 +154,9 @@ public:
 
     uint32_t get_n_kv() const;
 
+    // TODO: temporary
+    bool get_supports_set_rows() const;
+
     // get views of the current state of the cache
     ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
     ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
@@ -227,7 +230,7 @@ private:
 
     // env: LLAMA_SET_ROWS (temporary)
     // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    int supports_set_rows = false;
+    bool supports_set_rows = false;
 
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
@@ -270,15 +273,13 @@ private:
                           float   freq_base,
                           float   freq_scale) const;
 
-    llm_graph_result_ptr build_graph_shift(
-            const llama_cparams & cparams,
-                   ggml_context * ctx,
-                    ggml_cgraph * gf) const;
+    ggml_cgraph * build_graph_shift(
+               llm_graph_result * res,
+                  llama_context * lctx) const;
 
-    llm_graph_result_ptr build_graph_defrag(
-            const llama_cparams & cparams,
-                   ggml_context * ctx,
-                    ggml_cgraph * gf,
+    ggml_cgraph * build_graph_defrag(
+               llm_graph_result * res,
+                  llama_context * lctx,
               const defrag_info & dinfo) const;
 
     struct cell_ranges_t {
@@ -340,6 +341,9 @@ public:
 
     uint32_t get_n_kv() const;
 
+    // TODO: temporary
+    bool get_supports_set_rows() const;
+
     // get views of the current state of the cache
     ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;

src/llama-memory-hybrid.cpp

@@ -38,9 +38,9 @@ llama_memory_hybrid::llama_memory_hybrid(
         type_v,
         v_trans,
         offload,
+        1,
         kv_size,
         n_seq_max,
-        1,
         n_pad,
         n_swa,
         swa_type

src/llama-memory-recurrent.cpp

@@ -446,7 +446,7 @@ bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) {
     // A slot should be always be contiguous.
 
     // can only process batches with an equal number of new tokens in each sequence
-    GGML_ASSERT(ubatch.equal_seqs);
+    GGML_ASSERT(ubatch.equal_seqs());
 
     int32_t min = size - 1;
     int32_t max = 0;
@@ -768,6 +768,8 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
     // Iterate and write all the keys first, each row is a cell
     // Get whole range at a time
     for (uint32_t il = 0; il < n_layer; ++il) {
+        // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
+        if (r_l[il] == nullptr) continue;
 
         // Write key type
         const int32_t r_type_i = (int32_t)r_l[il]->type;
@@ -787,6 +789,8 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
 
     if (!s_trans) {
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
+            if (s_l[il] == nullptr) continue;
 
             // Write value type
             const int32_t s_type_i = (int32_t)s_l[il]->type;
@@ -807,6 +811,9 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
         // When v is transposed, we also need the element size and get the element ranges from each row
         const uint32_t mem_size = size;
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
+            if (s_l[il] == nullptr) continue;
+
             const uint32_t n_embd_s = hparams.n_embd_s();
 
             // Write value type
@@ -951,6 +958,8 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
     // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block
     for (uint32_t il = 0; il < n_layer; ++il) {
+        // skip null layers
+        if (r_l[il] == nullptr) continue;
 
         // Read type of key
         int32_t r_type_i_ref;
@@ -978,11 +987,14 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
     if (!s_trans) {
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers
+            if (s_l[il] == nullptr) continue;
 
             // Read type of value
             int32_t s_type_i_ref;
             io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
             const int32_t s_type_i = (int32_t)s_l[il]->type;
+
             if (s_type_i != s_type_i_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched s type (%d != %d, layer %d)\n", __func__, s_type_i, s_type_i_ref, il);
                 return false;
@@ -1005,6 +1017,9 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
     } else {
         // For each layer, read the values for each cell (transposed)
         for (uint32_t il = 0; il < n_layer; ++il) {
+            // skip null layers
+            if (s_l[il] == nullptr) continue;
+
             const uint32_t n_embd_s = hparams.n_embd_s();
 
             // Read type of value

src/llama-model.h

@@ -99,8 +99,10 @@ enum llm_type {
     LLM_TYPE_17B_16E, // llama4 Scout
     LLM_TYPE_17B_128E, // llama4 Maverick
     LLM_TYPE_A13B,
+    LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
     LLM_TYPE_235B_A22B,
+    LLM_TYPE_300B_A47B, // Ernie MoE big
     LLM_TYPE_E2B,
     LLM_TYPE_E4B,
 };
@@ -452,10 +454,7 @@ struct llama_model {
     llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const;
 
     // TODO: move this to new llm_arch_model_i interface
-    llm_graph_result_ptr build_graph(
-            const llm_graph_params & params,
-                       ggml_cgraph * gf,
-                    llm_graph_type   type) const;
+    ggml_cgraph * build_graph(const llm_graph_params & params) const;
 
 private:
     struct impl;

src/llama-vocab.cpp

@@ -1925,6 +1925,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             } else if (
                 tokenizer_pre == "exaone") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_EXAONE;
+            } else if (
+                tokenizer_pre == "exaone4") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                 tokenizer_pre == "chameleon") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;

tests/test-backend-ops.cpp

@@ -2353,9 +2353,12 @@ struct test_bin_bcast : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const std::array<int, 4> nr;
+    int nf; // number of fused ops, nf == 1 -> single op (no fusion)
+
+    bool run_whole_graph() override { return true; }
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, nr);
+        return VARS_TO_STR4(type, ne, nr, nf);
     }
 
     size_t op_size(ggml_tensor * t) override {
@@ -2364,24 +2367,35 @@ struct test_bin_bcast : public test_case {
 
     test_bin_bcast(op_t op, ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 1, 1},
-            std::array<int, 4> nr = {1, 2, 1, 1})
-        : op(op), type(type), ne(ne), nr(nr) {}
+            std::array<int, 4> nr = {1, 2, 1, 1},
+            int nf = 1)
+        : op(op), type(type), ne(ne), nr(nr), nf(nf) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
+        GGML_ASSERT(nf <= 8);
+
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0]*nr[0], ne[1]*nr[1], ne[2]*nr[2], ne[3]*nr[3]);
         ggml_set_name(a, "a");
 
-        ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_name(b, "b");
-
-        // The backward pass supports broadcasting only for GGML_ADD:
-        const bool grad_supported = op == ggml_add || ggml_are_same_shape(a, b);
-        if (grad_supported) {
-            ggml_set_param(a);
-            ggml_set_param(b);
+        ggml_tensor * b[8];
+        for (int i = 0; i < nf; ++i) {
+            b[i] = ggml_new_tensor(ctx, type, 4, ne.data());
+            ggml_set_name(b[i], (std::string("b") + std::to_string(i)).c_str());
+        }
+
+        // The backward pass supports broadcasting only for GGML_ADD:
+        const bool grad_supported = op == ggml_add && ggml_are_same_shape(a, b[0]) && nf == 1;
+        if (grad_supported) {
+            ggml_set_param(a);
+            ggml_set_param(b[0]);
+        }
+
+        ggml_tensor * out = a;
+
+        for (int i = 0; i < nf; ++i) {
+            out = op(ctx, out, b[i]);
         }
 
-        ggml_tensor * out = op(ctx, a, b);
         ggml_set_name(out, "out");
 
         return out;
@@ -2622,39 +2636,46 @@ struct test_rms_norm_back : public test_case {
     }
 };
 
-// GGML_OP_RMS_NORM + GGML_OP_MUL
-struct test_rms_norm_mul : public test_case {
+// GGML_OP_RMS_NORM + GGML_OP_MUL + GGML_OP_ADD
+struct test_rms_norm_mul_add : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const float eps;
+    const bool broadcast;
 
     std::string op_desc(ggml_tensor * t) override {
         GGML_UNUSED(t);
-        return "RMS_NORM_MUL";
+        return "RMS_NORM_MUL_ADD";
     }
 
     bool run_whole_graph() override { return true; }
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, eps);
+        return VARS_TO_STR4(type, ne, eps, broadcast);
     }
 
-    test_rms_norm_mul(ggml_type type = GGML_TYPE_F32,
+    test_rms_norm_mul_add(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {64, 5, 4, 3},
-            float eps = 1e-6f)
-        : type(type), ne(ne), eps(eps) {}
+            float eps = 1e-6f, bool broadcast = false)
+        : type(type), ne(ne), eps(eps), broadcast(broadcast) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        std::array<int64_t, 4> broadcast_dims = {ne[0]*2, ne[1]*3, ne[2]*3, ne[3]*4};
+
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, broadcast ? broadcast_dims.data() : ne.data());
         ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * c = ggml_new_tensor(ctx, type, 4, ne.data());
+
         ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_set_param(b);
         ggml_set_name(b, "b");
+        ggml_set_param(c);
+        ggml_set_name(c, "c");
 
-        // Use a and b early, so we don't end up with an OP_NONE between rms_norm and mul
-        a = ggml_add(ctx, a, b);
-        ggml_tensor * out = ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), b);
+        // Use a, b and c early, so we don't end up with an OP_NONE between rms_norm and mul
+        a = ggml_add(ctx, ggml_add(ctx, a, b), c);
+        ggml_tensor * out = ggml_add(ctx, ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), b), c);
         ggml_set_name(out, "out");
 
         return out;
@@ -3682,6 +3703,93 @@ struct test_im2col : public test_case {
     }
 };
 
+// CONV_2D
+struct test_conv_2d : public test_case {
+    const std::array<int64_t, 4> ne_input;
+    const std::array<int64_t, 4> ne_kernel;
+    const int                    stride0;
+    const int                    stride1;
+    const int                    padding0;
+    const int                    padding1;
+    const int                    dilation0;
+    const int                    dilation1;
+    // Whether the inputs are contiguous in the channel dim or the width dim
+    const bool                   cwhn;
+
+    // If true, the direct CONV_2D will be used in the graph, otherwise it
+    // uses ggml_conv_2d:
+    // * if the program is called with -o CONV_2D_DIRECT_IMPL, the
+    // CONV_2D graph will be built, while
+    // * if the program is called with -o CONV_2D_INDIRECT_IMPL, the
+    // IM2COL -> MUL_MM graph will be built.
+
+    std::string vars() override {
+        return VARS_TO_STR9(ne_input, ne_kernel, stride0, stride1, padding0, padding1, dilation0, dilation1, cwhn);
+    }
+
+    uint64_t op_flops(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        // Just counting matmul costs:
+        // KxCRS @ CRSxNPQ = KxNPQ --> KxNPQx(CRS+CRS-1) flops
+
+        // Copied from ggml.c: int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d)
+        auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+            return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
+        };
+
+        int64_t W    = ne_input[0];
+        int64_t H    = ne_input[1];
+        int64_t KW   = ne_kernel[0];
+        int64_t KH   = ne_kernel[1];
+        int64_t Cin  = ne_kernel[2];
+        int64_t Cout = ne_kernel[3];
+        int64_t N    = ne_input[3];
+        int64_t OH   = calc_conv_output_size(H, KH, stride0, padding0, dilation0);
+        int64_t OW   = calc_conv_output_size(W, KW, stride0, padding0, dilation0);
+
+        int64_t K   = Cout;
+        int64_t CRS = Cin * KH * KW;
+        int64_t NPQ = N * OH * OW;
+
+        return K * NPQ * (2 * CRS - 1);
+    }
+
+    test_conv_2d(std::array<int64_t, 4> ne_input  = { 64, 64, 16, 1 },
+                 std::array<int64_t, 4> ne_kernel = { 3, 3, 1, 16 }, int stride0 = 1, int stride1 = 1, int padding0 = 0,
+                 int padding1 = 0, int dilation0 = 1, int dilation1 = 1, bool cwhn = false) :
+        ne_input(ne_input),
+        ne_kernel(ne_kernel),
+        stride0(stride0),
+        stride1(stride1),
+        padding0(padding0),
+        padding1(padding1),
+        dilation0(dilation0),
+        dilation1(dilation1),
+        cwhn(cwhn) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data());
+        ggml_set_name(input, "input");
+
+        ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_kernel.data());
+        ggml_set_name(kernel, "kernel");
+
+        if (cwhn) {
+            // change memory layout to channel-most-contiguous (CWHN),
+            // then permute it back so NE matches the original input
+            input  = ggml_cont(ctx, ggml_permute(ctx, input, 1, 2, 0, 3));
+            input  = ggml_permute(ctx, input, 2, 0, 1, 3);
+            kernel = ggml_cont(ctx, ggml_permute(ctx, kernel, 2, 3, 1, 0));
+            kernel = ggml_permute(ctx, kernel, 3, 2, 0, 1);
+        }
+
+        ggml_tensor * out =
+            ggml_conv_2d_direct(ctx, kernel, input, stride0, stride1, padding0, padding1, dilation0, dilation1);
+        ggml_set_name(out, "out");
+        return out;
+    }
+};
+
 // GGML_OP_CONV_2D_DW
 struct test_conv_2d_dw : public test_case {
     const std::array<int64_t, 4> ne_input;
@@ -4258,26 +4366,32 @@ struct test_flash_attn_ext : public test_case {
         const int64_t hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_KV));
         const int64_t hsv_padded = GGML_PAD(hsv, ggml_blck_size(type_KV));
 
-        auto const &create_permuted = [&](ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) -> ggml_tensor * {
+        auto const &create_permuted = [&](ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, bool is_view) -> ggml_tensor * {
             int64_t ne[4] = {ne0, ne1, ne2, ne3};
             int64_t ne_perm[4];
             for (int i = 0; i < 4; ++i) {
                 ne_perm[permute[i]] = ne[i];
             }
-            ggml_tensor * t = ggml_new_tensor_4d(ctx, type, ne_perm[0], ne_perm[1], ne_perm[2], ne_perm[3]);
+            ggml_tensor * t;
+            if (is_view) {
+                ggml_tensor * t0 = ggml_new_tensor_4d(ctx, type, ne_perm[0], 2*ne_perm[1], ne_perm[2], ne_perm[3]);
+                t = ggml_view_4d(ctx, t0, ne_perm[0], ne_perm[1], ne_perm[2], ne_perm[3], t0->nb[1], t0->nb[2], t0->nb[3], 0);
+            } else {
+                t = ggml_new_tensor_4d(ctx, type, ne_perm[0], ne_perm[1], ne_perm[2], ne_perm[3]);
+            }
             if (permute != std::array<int32_t, 4>{0, 1, 2, 3}) {
                 t = ggml_permute(ctx, t, permute[0], permute[1], permute[2], permute[3]);
             }
             return t;
         };
 
-        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr23[0], nr23[1]);
+        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr23[0], nr23[1], false);
         ggml_set_name(q, "q");
 
-        ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,         nr23[1]);
+        ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,         nr23[1], true); // the K tensor is usually a view of the K cache
         ggml_set_name(k, "k");
 
-        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,         nr23[1]);
+        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,         nr23[1], true); // the V tensor is usually a view of the V cache
         ggml_set_name(v, "v");
 
         ggml_tensor * m = nullptr;
@@ -4989,6 +5103,81 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2048}, {3, 3, 2, 2048}, 1, 1, 1, 1, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 1, 2560}, {3, 3, 1, 2560}, 1, 1, 1, 1, 1, 1, true));
     test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {12, 12, 2, 2560}, {3, 3, 2, 2560}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {5, 5, 1, 32}, {3, 4, 1, 32}, 1, 1, 0, 0, 1, 1, true));
+
+// Conv_2D test cases
+#ifdef DETAILED_TESTS
+    // Probably we do not have enough time to execute these in the pipeline.
+    uint32_t iwh_idx  = 0;
+    uint32_t kwh_idx  = 1;
+    uint32_t Cout_idx = 2;
+    uint32_t Cin_idx  = 3;
+    uint32_t B_idx    = 4;
+
+    std::vector<std::array<int, 5>> cases = {
+  //{IWH, KWH, Cout, Cin, B}
+  // K=CRS=NPQ=4096 conv_2d matmul performance
+        {19,   4, 4096, 256, 16},
+ // K=128, CRS=128, NPQ=4096
+        { 19,  4, 128,  8,   16},
+ // K=130, CRS=128, NPQ=4096
+        { 19,  4, 130,  8,   16},
+ // Edge case: K x CRS is small
+        { 19,  2, 4,    4,   16},
+ // A ConvNet's first layer
+        { 224, 3, 8,    3,   1 },
+ // A ConvNet's first layer with 2x2 convolution, and 1 channel
+        { 224, 2, 8,    1,   1 },
+ // A ConvNet's first layer with 2x2 convolution, and 1 channel, several images in the batch
+        { 224, 2, 8,    1,   8 },
+ // A middle layer of a ConvNet
+        { 58,  3, 64,   32,  1 },
+ // A middle layer of a ConvNet, several images in the batch
+        { 58,  3, 64,   32,  8 },
+ // A deep layer of a ConvNet, several images in the batch
+        { 16,  3, 256,  128, 8 }
+    };
+
+    for (auto act_case : cases) {
+        test_cases.emplace_back(new test_conv_2d(
+            { act_case[iwh_idx], act_case[iwh_idx], act_case[Cin_idx], act_case[B_idx] },
+            { act_case[kwh_idx], act_case[kwh_idx], act_case[Cin_idx], act_case[Cout_idx] }, 1, 1, 0, 0, 1, 1, false));
+    }
+#endif
+
+    // CONV_2D:
+    auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+        return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
+    };
+
+    //uint32_t s0 = 3;
+    uint32_t s1 = 5;
+    uint32_t p0 = 5;
+    //uint32_t p1 = 2;
+    uint32_t d0 = 2;
+    uint32_t d1 = 4;
+
+    for (uint32_t s0 : { 1, 3 }) {
+        for (uint32_t p1 : { 2, 5 }) {
+            for (uint32_t Cin : { 1, 25 }) {
+                for (uint32_t Cout : { 1, 12 }) {
+                    for (uint32_t KH : { 1, 2, 3, 11 }) {
+                        for (uint32_t KW : { 1, 2, 3, 11 }) {
+                            for (uint32_t H : { 1, 133 }) {
+                                for (uint32_t W : { 1, 141 }) {
+                                    if (calc_conv_output_size(W, KW, s0, p0, d0) > 0 &&
+                                        calc_conv_output_size(H, KH, s1, p1, d1) > 0) {
+                                        test_cases.emplace_back(new test_conv_2d(
+                                            { W, H, Cin, 2 }, { KW, KH, Cin, Cout }, s0, s1, p0, p1, d0, d1, false));
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
 
     // sycl backend will limit task global_range < MAX_INT
     // test cases for 2D im2col with large input W and H (occurs in stable-diffusion)
@@ -5151,6 +5340,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         //add_test_bin_bcast(type, {3, 3, 2560, 1280}, {2, 1, 1, 1});
     }
 
+    // fusion
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {2, 1, 1, 1}, 2));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {16, 5, 4, 3}, {1, 2, 1, 1}, 3));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 1}, 4));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {16, 5, 4, 3}, {1, 1, 1, 2}, 5));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 2}, 6));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 2, 2}, 7));
+    test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {16, 5, 4, 3}, {2, 2, 2, 2}, 8));
+
     test_cases.emplace_back(new test_add1());
     test_cases.emplace_back(new test_scale());
     test_cases.emplace_back(new test_scale(GGML_TYPE_F32, {10, 10, 10, 10}, 2.0f, 1.0f));
@@ -5165,7 +5363,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_l2_norm      (GGML_TYPE_F32, {64, 5, 4, 3}, eps));
     }
     for (float eps : {0.0f, 1e-6f, 1e-4f, 1e-1f, 1.0f}) {
-        test_cases.emplace_back(new test_rms_norm_mul(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+        test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+        test_cases.emplace_back(new test_rms_norm_mul_add(GGML_TYPE_F32, {64, 5, 4, 3}, eps, true));
     }
 
     test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));
@@ -5584,6 +5783,43 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     std::vector<std::unique_ptr<test_case>> test_cases;
 
+    // Conv2d: K=CRS=NPQ=4096 matmul performance
+    uint32_t                        iwh_idx  = 0;
+    uint32_t                        kwh_idx  = 1;
+    uint32_t                        Cout_idx = 2;
+    uint32_t                        Cin_idx  = 3;
+    uint32_t                        B_idx    = 4;
+    std::vector<std::array<int, 5>> cases    = {
+  //{IWH, KWH, Cout, Cin, B}
+  // K=CRS=NPQ=4096 conv2d matmul performance
+        {19,   4, 4096, 256, 16},
+ // K=128, CRS=128, NPQ=4096
+        { 19,  4, 128,  8,   16},
+ // K=130, CRS=128, NPQ=4096
+        { 19,  4, 130,  8,   16},
+ // Edge case: K x CRS is small
+        { 19,  2, 4,    4,   16},
+ // A ConvNet's first layer
+        { 224, 3, 8,    3,   1 },
+ // A ConvNet's first layer with 2x2 convolution, and 1 channel
+        { 224, 2, 8,    1,   1 },
+ // A ConvNet's first layer with 2x2 convolution, and 1 channel, several images in the batch
+        { 224, 2, 8,    1,   8 },
+ // A middle layer of a ConvNet
+        { 58,  3, 64,   32,  1 },
+ // A middle layer of a ConvNet, several images in the batch
+        { 58,  3, 64,   32,  8 },
+ // A deep layer of a ConvNet, several images in the batch
+        { 16,  3, 512,  128, 8 },
+    };
+
+    for (auto act_case : cases) {
+        // Direct CONV_2D
+        test_cases.emplace_back(new test_conv_2d(
+            { act_case[iwh_idx], act_case[iwh_idx], act_case[Cin_idx], act_case[B_idx] },
+            { act_case[kwh_idx], act_case[kwh_idx], act_case[Cin_idx], act_case[Cout_idx] }, 1, 1, 0, 0, 1, 1, false));
+    }
+
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1,   1, 1, 1}));
     test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 512, 1, 1}));
 

tools/imatrix/README.md

@@ -1,33 +1,92 @@
 # llama.cpp/tools/imatrix
 
 Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models.
-More information is available here: https://github.com/ggml-org/llama.cpp/pull/4861
+More information is available in <https://github.com/ggml-org/llama.cpp/pull/4861>.
 
 ## Usage
 
 ```
 ./llama-imatrix \
-    -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \
-    [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \
-    [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]
+    -m model.gguf -f some-text.txt [-o imatrix.gguf] [--no-ppl] \
+    [--process-output] [--chunk 123] [--save-frequency 0] [--output-frequency 10] \
+    [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...] [--parse-special] \
+    [--show-statistics] [...]
 ```
 
-Here `-m` with a model name and `-f` with a file containing training data (such as e.g. `wiki.train.raw`) are mandatory.
+Here `-m | --model` with a model name and `-f | --file` with a file containing calibration data (such as e.g. `wiki.train.raw`) are mandatory.
 The parameters in square brackets are optional and have the following meaning:
-* `-o` (or `--output-file`) specifies the name of the file where the computed data will be stored. If missing `imatrix.dat` is used.
-* `--verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
-* `--output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
+
+* `-h | --help` shows usage information and exits.
+* `-lv | --verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
+* `-o | --output-file` specifies the name of the file where the computed data will be stored. If missing `imatrix.gguf` is used.
+* `-ofreq | --output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
 * `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
-* `--process-output` specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
+* `--process-output` specifies if data will be collected for the `output.weight` tensor. Typically, it is better not to utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
+* `--in-file` one or more existing imatrix files to load and combine. Useful for merging files from multiple runs/datasets.
+* `--parse-special` enables parsing of special tokens (e.g., `<|im_start|>` in some models). Useful for models with custom tokenizers.
+* `--chunk | --from-chunk` to skip the first `n` chunks of tokens from the input data. Useful for resuming or skipping initial low-quality data.
+* `--chunks` maximum number of chunks to process. Default is -1 for all available chunks.
+* `--no-ppl` disables the calculation of perplexity for the processed chunks. Useful if you want to speed up the processing and do not care about perplexity.
+* `--show-statistics` displays imatrix file's statistics.
 
-For faster computation, make sure to use GPU offloading via the `-ngl` argument
+For faster computation, make sure to use GPU offloading via the `-ngl | --n-gpu-layers` argument.
 
-## Example
+Recent versions of `llama-imatrix` store data in GGUF format by default. For the legacy format, use an extension other than `.gguf` when saving the output file. More information is available in <https://github.com/ggml-org/llama.cpp/pull/9400>.
+
+## Examples
 
 ```bash
-# generate importance matrix (imatrix.dat)
-./llama-imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
+# generate importance matrix using default filename (imatrix.gguf), offloading 99 layers to GPU
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt -ngl 99
 
 # use the imatrix to perform a Q4_K_M quantization
-./llama-quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
+./llama-quantize --imatrix imatrix.gguf ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
 ```
+
+```bash
+# generate and save the imatrix using legacy format
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt -o imatrix-legcy-format.dat -ngl 99
+```
+
+```bash
+# covert legacy (binary) imatrix format to new (GGUF) format
+./llama-imatrix --in-file imatrix-legacy-format.dat -o imatrix-new-format.gguf
+```
+
+```bash
+# combine existing imatrices
+./llama-imatrix --in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf -o imatrix-combined.gguf
+```
+
+```bash
+# skip first 5 chunks, save intermediates every 20 chunks and snapshots every 50, parsing special tokens
+./llama-imatrix -m ggml-model-f16.gguf -f calibration-data.txt --chunk 5 --output-frequency 20 --save-frequency 50 --parse-special
+```
+
+```bash
+# analyse imatrix file and display summary statistics instead of running inference
+./llama-imatrix --in-file imatrix.gguf --show-statistics
+```
+
+`--show-statistics` will display the following statistics:
+
+#### Per tensor
+
+* Σ(Act²): sum of all squared activations (the importance scores)
+* Min & Max: minimum and maximum squared activations values
+* μ & σ: Squared activations' mean and standard deviation
+* % Active: proportion of elements whose average squared activation exceeds a small threshold (1e-5). Helpful to determine how alive/dormant the tensor is during inference
+* N: number of squared activations
+* Entropy: entropy of the squared activation distribution, in bits (standard Shannon entropy measurement) $S = -\sum_{i=1}^N p_i \log_2 p_i$
+* E (norm): Normalized entropy. $E(norm)=\frac{-\sum_{i=1}^N p_i \log_2 p_i}{log_2 N}$. These two metrics can be used to determine how well a prompt "exercises" the model's capabilities
+* ZD Score: z-score distribution as described in _3.1 Layer Importance Scores_ of [Layer-Wise Quantization](https://arxiv.org/abs/2406.17415)
+* CosSim: cosine similarity with respect to the previous layer's tensor. Useful to determine how similar the squared activations of the current layer are to the previous layer's squared activations.
+
+#### Per layer
+
+Weighted averages of Σ(Act²), ZD Score and CosSim are also calculated.
+
+#### Important note on the computed Statistics
+
+When using these statistics, please note that they are computed on the squared activations, **not on the actual (raw) activations**.
+Whilst the results are still useful, they're less realiable than using the raw values, and in the case of the cosine similarity, could be misleading if the tensor contains opposite vectors.

tools/main/main.cpp

@@ -785,14 +785,17 @@ int main(int argc, char ** argv) {
                 }
 
                 // check for reverse prompt using special tokens
-                llama_token last_token = common_sampler_last(smpl);
-                for (auto token : antiprompt_token) {
-                    if (token == last_token) {
-                        if (params.interactive) {
-                            is_interacting = true;
+                // avoid calling common_sampler_last() if last_output is empty
+                if (!last_output.empty()) {
+                    llama_token last_token = common_sampler_last(smpl);
+                    for (auto token : antiprompt_token) {
+                        if (token == last_token) {
+                            if (params.interactive) {
+                                is_interacting = true;
+                            }
+                            is_antiprompt = true;
+                            break;
                         }
-                        is_antiprompt = true;
-                        break;
                     }
                 }
 

tools/mtmd/clip.cpp

@@ -367,8 +367,8 @@ struct clip_ctx {
     std::vector<ggml_backend_t> backend_ptrs;
     std::vector<ggml_backend_buffer_type_t> backend_buft;
 
-    ggml_backend_t backend;
-    ggml_backend_t backend_cpu;
+    ggml_backend_t backend = nullptr;
+    ggml_backend_t backend_cpu = nullptr;
     ggml_backend_buffer_ptr buf;
 
     int max_nodes = 8192;
@@ -384,9 +384,18 @@ struct clip_ctx {
         if (!backend_cpu) {
             throw std::runtime_error("failed to initialize CPU backend");
         }
-        backend = ctx_params.use_gpu
-                    ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr)
-                    : nullptr;
+        if (ctx_params.use_gpu) {
+            auto backend_name = std::getenv("MTMD_BACKEND_DEVICE");
+            if (backend_name != nullptr) {
+                backend = ggml_backend_init_by_name(backend_name, nullptr);
+                if (!backend) {
+                    LOG_WRN("%s: Warning: Failed to initialize \"%s\" backend, falling back to default GPU backend\n", __func__, backend_name);
+                }
+            }
+            if (!backend) {
+                backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
+            }
+        }
 
         if (backend) {
             LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend));

tools/quantize/quantize.cpp

@@ -1,11 +1,13 @@
 #include "common.h"
 #include "llama.h"
+#include "gguf.h"
 
 #include <cstdio>
 #include <cstring>
 #include <vector>
 #include <string>
 #include <unordered_map>
+#include <map>
 #include <fstream>
 #include <cmath>
 #include <cctype>
@@ -68,6 +70,11 @@ static const char * const LLM_KV_QUANTIZE_IMATRIX_DATASET    = "quantize.imatrix
 static const char * const LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES  = "quantize.imatrix.entries_count";
 static const char * const LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS   = "quantize.imatrix.chunks_count";
 
+// TODO: share with imatrix.cpp
+static const char * const LLM_KV_IMATRIX_DATASETS    = "imatrix.datasets";
+static const char * const LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count";
+static const char * const LLM_KV_IMATRIX_CHUNK_SIZE  = "imatrix.chunk_size";
+
 static bool striequals(const char * a, const char * b) {
     while (*a && *b) {
         if (std::tolower(*a) != std::tolower(*b)) {
@@ -84,7 +91,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
     for (auto ch : ftype_str_in) {
         ftype_str.push_back(std::toupper(ch));
     }
-    for (auto & it : QUANT_OPTIONS) {
+    for (const auto & it : QUANT_OPTIONS) {
         if (striequals(it.name.c_str(), ftype_str.c_str())) {
             ftype = it.ftype;
             ftype_str_out = it.name;
@@ -93,7 +100,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
     }
     try {
         int ftype_int = std::stoi(ftype_str);
-        for (auto & it : QUANT_OPTIONS) {
+        for (const auto & it : QUANT_OPTIONS) {
             if (it.ftype == ftype_int) {
                 ftype = it.ftype;
                 ftype_str_out = it.name;
@@ -129,7 +136,7 @@ static void usage(const char * executable) {
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
     printf("Note: --include-weights and --exclude-weights cannot be used together\n");
     printf("\nAllowed quantization types:\n");
-    for (auto & it : QUANT_OPTIONS) {
+    for (const auto & it : QUANT_OPTIONS) {
         if (it.name != "COPY") {
             printf("  %2d  or  ", it.ftype);
         } else {
@@ -140,7 +147,7 @@ static void usage(const char * executable) {
     exit(1);
 }
 
-static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_dataset, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
+static int load_legacy_imatrix(const std::string & imatrix_file, std::vector<std::string> & imatrix_datasets, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
     std::ifstream in(imatrix_file.c_str(), std::ios::binary);
     if (!in) {
         printf("%s: failed to open %s\n",__func__, imatrix_file.c_str());
@@ -180,7 +187,9 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
             exit(1);
         }
         if (ncall > 0) {
-            for (auto& v : e) v /= ncall;
+            for (auto & v : e) {
+                v /= ncall;
+            }
         }
 
         if (getenv("LLAMA_TRACE")) {
@@ -188,7 +197,7 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
         }
     }
 
-    // latest imatrix version contains the dataset filename at the end of the file
+    // latest legacy imatrix version contains the dataset filename at the end of the file
     int m_last_call = 0;
     if (in.peek() != EOF) {
         in.read((char *)&m_last_call, sizeof(m_last_call));
@@ -196,15 +205,130 @@ static int load_imatrix(const std::string & imatrix_file, std::string & imatrix_
         in.read((char *)&dataset_len, sizeof(dataset_len));
         std::vector<char> dataset_as_vec(dataset_len);
         in.read(dataset_as_vec.data(), dataset_len);
-        imatrix_dataset.assign(dataset_as_vec.begin(), dataset_as_vec.end());
-        printf("%s: imatrix dataset='%s'\n", __func__, imatrix_dataset.c_str());
+        imatrix_datasets.resize(1);
+        imatrix_datasets[0].assign(dataset_as_vec.begin(), dataset_as_vec.end());
+        printf("%s: imatrix dataset='%s'\n", __func__, imatrix_datasets[0].c_str());
     }
     printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_call);
     return m_last_call;
 }
 
+static int load_imatrix(const std::string & imatrix_file, std::vector<std::string> & imatrix_datasets, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
+
+    struct ggml_context * ctx = nullptr;
+    struct gguf_init_params meta_gguf_params = {
+        /* .no_alloc = */ false, // the data is needed
+        /* .ctx      = */ &ctx,
+    };
+    struct gguf_context * ctx_gguf = gguf_init_from_file(imatrix_file.c_str(), meta_gguf_params);
+    if (!ctx_gguf) {
+        fprintf(stderr, "%s: imatrix file '%s' is using old format\n", __func__, imatrix_file.c_str());
+        return load_legacy_imatrix(imatrix_file, imatrix_datasets, imatrix_data);
+    }
+    const int32_t n_entries = gguf_get_n_tensors(ctx_gguf);
+    if (n_entries < 1) {
+        fprintf(stderr, "%s: no data in file %s\n", __func__, imatrix_file.c_str());
+        gguf_free(ctx_gguf);
+        ggml_free(ctx);
+        exit(1);
+    }
+
+    const int dataset_idx     = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_DATASETS);
+    const int chunk_count_idx = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_COUNT);
+    const int chunk_size_idx  = gguf_find_key(ctx_gguf, LLM_KV_IMATRIX_CHUNK_SIZE);
+    if (dataset_idx < 0 || chunk_count_idx < 0 || chunk_size_idx < 0) {
+        fprintf(stderr, "%s: missing imatrix metadata in file %s\n", __func__, imatrix_file.c_str());
+        gguf_free(ctx_gguf);
+        ggml_free(ctx);
+        exit(1);
+    }
+
+    const uint32_t chunk_size = gguf_get_val_u32(ctx_gguf, chunk_size_idx);
+
+    const std::string sums_suffix{ ".in_sum2" };
+    const std::string counts_suffix{ ".counts" };
+
+    // Using an ordered map to get a deterministic iteration order.
+    std::map<std::string, std::pair<struct ggml_tensor *, struct ggml_tensor *>> sums_counts_for;
+
+    for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur; cur = ggml_get_next_tensor(ctx, cur)) {
+        std::string name = cur->name;
+
+        if (name.empty()) { continue; }
+
+        if (string_remove_suffix(name, sums_suffix)) {
+            // in_sum2
+            sums_counts_for[std::move(name)].first = cur;
+        } else if (string_remove_suffix(name, counts_suffix)) {
+            // counts
+            sums_counts_for[std::move(name)].second = cur;
+        } else {
+            // ignore other tensors
+        }
+    }
+
+    for (const auto & sc : sums_counts_for) {
+        const        std::string & name   = sc.first;
+        const struct ggml_tensor * sums   = sc.second.first;
+        const struct ggml_tensor * counts = sc.second.second;
+
+        if (!sums || !counts) {
+            fprintf(stderr, "%s: mismatched sums and counts for %s\n", __func__, name.c_str());
+            gguf_free(ctx_gguf);
+            ggml_free(ctx);
+            exit(1);
+        }
+
+        const int64_t ne0 = sums->ne[0];
+        const int64_t ne1 = sums->ne[1];
+
+        auto & e = imatrix_data[name];
+        e.resize(ggml_nelements(sums));
+        float max_count = 0.0f;
+        for (int64_t j = 0; j < ne1; ++j) {
+            const float count = ((const float *) counts->data)[j];
+            if (count > 0.0f) {
+                for (int64_t i = 0; i < ne0; ++i) {
+                    e[j*ne0 + i] = ((const float *) sums->data)[j*ne0 + i] / count;
+                }
+            } else {
+                // Partial imatrix data, this tensor never got any input during calibration
+                for (int64_t i = 0; i < ne0; ++i) {
+                    e[j*ne0 + i] = 1;
+                }
+            }
+            if (count > max_count) {
+                max_count = count;
+            }
+        }
+        if (getenv("LLAMA_TRACE")) {
+            printf("%s: loaded data (size = %6d, n_tokens = %6d, n_chunks = %6d) for '%s'\n", __func__, int(e.size()), int(max_count), int(max_count / chunk_size), name.c_str());
+        }
+    }
+
+    int m_last_chunk = gguf_get_val_u32(ctx_gguf, chunk_count_idx);
+
+    int64_t n_datasets = gguf_get_arr_n(ctx_gguf, dataset_idx);
+    imatrix_datasets.reserve(n_datasets);
+    for (int64_t i = 0; i < n_datasets; ++i) {
+        imatrix_datasets.push_back(gguf_get_val_str(ctx_gguf, dataset_idx));
+    }
+    printf("%s: imatrix datasets=['%s'", __func__, imatrix_datasets[0].c_str());
+    for (size_t i = 1; i < imatrix_datasets.size(); ++i) {
+        printf(", '%s'", imatrix_datasets[i].c_str());
+    }
+    printf("]\n");
+
+    printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_chunk);
+
+    gguf_free(ctx_gguf);
+    ggml_free(ctx);
+
+    return m_last_chunk;
+}
+
 static int prepare_imatrix(const std::string & imatrix_file,
-        std::string & imatrix_dataset,
+        std::vector<std::string> & imatrix_dataset,
         const std::vector<std::string> & included_weights,
         const std::vector<std::string> & excluded_weights,
         std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
@@ -216,18 +340,21 @@ static int prepare_imatrix(const std::string & imatrix_file,
         return m_last_call;
     }
     if (!excluded_weights.empty()) {
-        for (auto& name : excluded_weights) {
-            for (auto it = imatrix_data.begin(); it != imatrix_data.end(); ) {
+        for (const auto & name : excluded_weights) {
+            for (auto it = imatrix_data.begin(); it != imatrix_data.end();) {
                 auto pos = it->first.find(name);
-                if (pos != std::string::npos) it = imatrix_data.erase(it);
-                else ++it;
+                if (pos != std::string::npos) {
+                    it = imatrix_data.erase(it);
+                } else {
+                    ++it;
+                }
             }
         }
     }
     if (!included_weights.empty()) {
         std::unordered_map<std::string, std::vector<float>> tmp;
-        for (auto& name : included_weights) {
-            for (auto& e : imatrix_data) {
+        for (const auto & name : included_weights) {
+            for (auto & e : imatrix_data) {
                 auto pos = e.first.find(name);
                 if (pos != std::string::npos) {
                     tmp.emplace(std::move(e));
@@ -396,9 +523,9 @@ int main(int argc, char ** argv) {
         usage(argv[0]);
     }
 
-    std::string imatrix_dataset;
+    std::vector<std::string> imatrix_datasets;
     std::unordered_map<std::string, std::vector<float>> imatrix_data;
-    int m_last_call = prepare_imatrix(imatrix_file, imatrix_dataset, included_weights, excluded_weights, imatrix_data);
+    int m_last_call = prepare_imatrix(imatrix_file, imatrix_datasets, included_weights, excluded_weights, imatrix_data);
     if (!imatrix_data.empty()) {
         params.imatrix = &imatrix_data;
         {
@@ -409,11 +536,12 @@ int main(int argc, char ** argv) {
             kvo.val_str[127] = '\0';
             kv_overrides.emplace_back(std::move(kvo));
         }
-        if (!imatrix_dataset.empty()) {
+        if (!imatrix_datasets.empty()) {
             llama_model_kv_override kvo;
+            // TODO: list multiple datasets when there are more than one
             std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_DATASET);
             kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
-            strncpy(kvo.val_str, imatrix_dataset.c_str(), 127);
+            strncpy(kvo.val_str, imatrix_datasets[0].c_str(), 127);
             kvo.val_str[127] = '\0';
             kv_overrides.emplace_back(std::move(kvo));
         }

tools/server/README.md

@@ -575,6 +575,8 @@ These words will not be included in the completion, so make sure to add them to
 
 `add_special`: (Optional) Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
 
+`parse_special`: (Optional) Boolean indicating if special tokens should be tokenized. When `false` special tokens are treated as plaintext.  Default: `true`
+
 `with_pieces`: (Optional) Boolean indicating whether to return token pieces along with IDs.  Default: `false`
 
 **Response:**

tools/server/server.cpp

@@ -253,6 +253,7 @@ struct server_task {
         defaults.sampling    = params_base.sampling;
         defaults.speculative = params_base.speculative;
         defaults.n_keep      = params_base.n_keep;
+        defaults.antiprompt  = params_base.antiprompt;
 
         // enabling this will output extra debug information in the HTTP responses from the server
         params.verbose           = params_base.verbosity > 9;
@@ -490,6 +491,10 @@ struct server_task {
                     }
                 }
             }
+            // set reverse prompt from cli args if not set in the request
+            if (params.antiprompt.empty()) {
+                params.antiprompt = defaults.antiprompt;
+            }
         }
 
         {
@@ -4516,9 +4521,10 @@ int main(int argc, char ** argv) {
         json tokens_response = json::array();
         if (body.count("content") != 0) {
             const bool add_special = json_value(body, "add_special", false);
+            const bool parse_special = json_value(body, "parse_special", true);
             const bool with_pieces = json_value(body, "with_pieces", false);
 
-            llama_tokens tokens = tokenize_mixed(ctx_server.vocab, body.at("content"), add_special, true);
+            llama_tokens tokens = tokenize_mixed(ctx_server.vocab, body.at("content"), add_special, parse_special);
 
             if (with_pieces) {
                 for (const auto& token : tokens) {