vulkan: fix rms_norm_mul to handle broadcasting dim0 (#14817)

Signed-off-by: Molly Sophia <mollysophia379@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

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

common/common.cpp

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

common/common.h

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

convert_hf_to_gguf.py

@@ -669,6 +669,36 @@ class TextModel(ModelBase):
         # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script
         #       or pull the latest version of the model from Huggingface
         #       don't edit the hashes manually!
+        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
+            res = "glm4"
+        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
+            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
+            res = "minerva-7b"
+        if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664":
+            # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct
+            res = "hunyuan"
+        if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
+            res = "falcon-h1"
+        if chkhsh == "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-1B-Base
+            res = "falcon-h1"
+        if chkhsh == "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-7B-Base
+            res = "falcon-h1"
+        if chkhsh == "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-34B-Base
+            res = "falcon-h1"
+        if chkhsh == "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890":
+            # ref: https://huggingface.co/moonshotai/Kimi-K2-Base
+            res = "kimi-k2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -804,42 +834,18 @@ class TextModel(ModelBase):
         if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
             # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
             res = "seed-coder"
-        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
-            res = "chatglm-bpe"
-        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
-            res = "chatglm-bpe"
-        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
-            res = "glm4"
-        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
-            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
-            res = "minerva-7b"
-        if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664":
-            # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct
-            res = "hunyuan"
         if chkhsh == "b0a6b1c0bd5998ebd9df08611efde34a4ff03faed45ae09c43e6b31ebd4b94cf":
             # ref: https://huggingface.co/skt/A.X-4.0
             res = "a.x-4.0"
-        if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
-            res = "falcon-h1"
-        if chkhsh == "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-1B-Base
-            res = "falcon-h1"
-        if chkhsh == "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-7B-Base
-            res = "falcon-h1"
-        if chkhsh == "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-34B-Base
-            res = "falcon-h1"
         if chkhsh == "f6791d196f87ce6b56a7d234be618e0d58f8cda3549416635b2bebcd22cd95c4":
             # ref: https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct
             res = "midm-2.0"
         if chkhsh == "169bf0296a13c4d9b7672313f749eb36501d931022de052aad6e36f2bf34dd51":
             # ref: https://huggingface.co/LiquidAI/LFM2-Tokenizer
             res = "lfm2"
+        if chkhsh == "2085e1638f6c377a0aa4ead21b27bb4cb941bf800df86ed391011769c1758dfb":
+            # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B
+            res = "exaone4"
 
         if res is None:
             logger.warning("\n")
@@ -2775,6 +2781,76 @@ class Qwen2Model(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("DreamModel")
+class DreamModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.DREAM
+
+    def get_vocab_base(self) -> tuple[list[str], list[int], str]:
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+
+        vocab_dict = tokenizer.get_vocab()
+        vocab_size = self.hparams.get("vocab_size", len(vocab_dict))
+        assert max(vocab_dict.values()) < vocab_size
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab_dict.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            elif reverse_vocab[i] in added_vocab:
+                tokens.append(reverse_vocab[i])
+                # Check if it's a special token - treat special tokens as CONTROL tokens
+                if hasattr(tokenizer, 'added_tokens_decoder') and i in tokenizer.added_tokens_decoder:
+                    if tokenizer.added_tokens_decoder[i].special:
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    # Fallback: treat all added vocab as control tokens for special tokens like <|im_start|>
+                    toktypes.append(gguf.TokenType.CONTROL)
+            else:
+                tokens.append(reverse_vocab[i])
+                toktypes.append(gguf.TokenType.NORMAL)
+
+        return tokens, toktypes, tokpre
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self._try_set_pooling_type()
+
+        # Dream models use non-causal attention for diffusion
+        self.gguf_writer.add_causal_attention(False)
+        # Handle RoPE scaling similar to Qwen2
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+        # Add Dream-specific parameters
+        mask_token_id = self.hparams.get("mask_token_id")
+        if mask_token_id is not None:
+            self.gguf_writer.add_mask_token_id(mask_token_id)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Dream model tensors should be mapped directly since it's the base model
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Ernie4_5_ForCausalLM")
 class Ernie4_5Model(TextModel):
     model_arch = gguf.MODEL_ARCH.ERNIE4_5
@@ -2788,7 +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.")
@@ -2821,6 +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",
@@ -3508,6 +3672,175 @@ class PlamoModel(TextModel):
         return [(new_name, data_torch)]
 
 
+@ModelBase.register("Plamo2ForCausalLM", "PLaMo2ForCausalLM")
+class Plamo2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.PLAMO2
+
+    def set_vocab(self):
+        # PLaMo 2 uses a custom tokenizer with a .jsonl file
+        # We need to handle this specially
+        tokenizer_jsonl_path = self.dir_model / "tokenizer.jsonl"
+        tokenizer_config_path = self.dir_model / "tokenizer_config.json"
+
+        if not tokenizer_jsonl_path.is_file():
+            raise FileNotFoundError(f"PLaMo 2 tokenizer file not found: {tokenizer_jsonl_path}")
+
+        # Load tokenizer config
+        with open(tokenizer_config_path, 'r', encoding='utf-8') as f:
+            tokenizer_config = json.load(f)
+
+        # Load tokens from JSONL file (actually a list format)
+        tokens = []
+        scores = []
+        toktypes = []
+
+        with open(tokenizer_jsonl_path, 'r', encoding='utf-8') as f:
+            for line_num, line in enumerate(f):
+                if line.strip():
+                    token_data = json.loads(line)
+                    # Format: [token, score, type, ?, ?, ?, ?]
+                    token = token_data[0].encode("utf-8")
+                    score = float(token_data[1])
+                    token_type_str = token_data[2] if len(token_data) > 2 else "NORMAL"
+
+                    tokens.append(token)
+                    scores.append(score)
+
+                    # Map token type strings to GGUF token types
+                    if token_type_str == "UNKNOWN":
+                        toktypes.append(gguf.TokenType.UNKNOWN)
+                    elif token_type_str == "CONTROL":
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    elif token_type_str == "BYTE":
+                        toktypes.append(gguf.TokenType.BYTE)
+                    else:
+                        # Check for PLaMo-2 special tokens
+                        token_str = token_data[0]
+                        if token_str.startswith("<|plamo:") and token_str.endswith("|>"):
+                            toktypes.append(gguf.TokenType.CONTROL)
+                        else:
+                            toktypes.append(gguf.TokenType.NORMAL)
+
+        vocab_size = self.hparams["vocab_size"]
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(gguf.TokenType.UNUSED)
+
+        # Use "plamo2" tokenizer type for PLaMo-2's custom Aho-Corasick tokenizer
+        self.gguf_writer.add_tokenizer_model("plamo2")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # Add special tokens from config
+        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] is not None:
+            token_id = tokens.index(tokenizer_config["bos_token"].encode("utf-8"))
+            self.gguf_writer.add_bos_token_id(token_id)
+        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] is not None:
+            token_id = tokens.index(tokenizer_config["eos_token"].encode("utf-8"))
+            self.gguf_writer.add_eos_token_id(token_id)
+        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] is not None:
+            token_id = tokens.index(tokenizer_config["pad_token"].encode("utf-8"))
+            self.gguf_writer.add_pad_token_id(token_id)
+        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] is not None:
+            token_id = tokens.index(tokenizer_config["sep_token"].encode("utf-8"))
+            self.gguf_writer.add_sep_token_id(token_id)
+        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] is not None:
+            token_id = tokens.index(tokenizer_config["unk_token"].encode("utf-8"))
+            self.gguf_writer.add_unk_token_id(token_id)
+
+        # Add <|plamo:op|> as EOT to ensure appropriate end of generation
+        self.gguf_writer.add_eot_token_id(4)
+
+        self.gguf_writer.add_add_space_prefix(False)
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+        block_count = hparams["num_hidden_layers"]
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+
+        # Which layers are Mamba layers
+        # PLaMo 2 uses mamba_step to indicate the pattern (e.g., 2 means every other layer)
+        # This logic matches modeling_plamo.py's is_mamba function
+        mamba_step = hparams.get("mamba_step", 2)
+        mamba_enabled = hparams.get("mamba_enabled", True)
+        mamba_layers = []
+
+        if mamba_enabled:
+            for i in range(block_count):
+                if block_count <= (mamba_step // 2):
+                    # use attention in last layer
+                    is_mamba = (i != block_count - 1)
+                else:
+                    is_mamba = (i % mamba_step) != (mamba_step // 2)
+                if is_mamba:
+                    mamba_layers.append(0)
+                else:
+                    mamba_layers.append(hparams.get("num_key_value_heads", 4))
+
+        if mamba_layers:
+            self.gguf_writer.add_head_count_kv(mamba_layers)
+
+        self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 2048))
+        self.gguf_writer.add_embedding_length(hparams.get("hidden_size", 4096))
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 32))
+        self.gguf_writer.add_layer_norm_rms_eps(hparams.get("rms_norm_eps", 1e-06))
+        self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1000000.0))
+
+        # Mamba parameters
+        self.gguf_writer.add_ssm_state_size(hparams.get("mamba_d_state", 64))
+        self.gguf_writer.add_ssm_conv_kernel(hparams.get("mamba_d_conv", 4))
+        self.gguf_writer.add_ssm_time_step_rank(hparams.get("mamba_num_heads", 64))
+        intermediate_size = hparams.get("mamba_num_heads", 64) * hparams.get("hidden_size_per_head", 128)
+        self.gguf_writer.add_ssm_inner_size(intermediate_size)
+        self.gguf_writer.add_ssm_group_count(0)
+
+        # MLP feed forward parameters (for attention layers)
+        self.gguf_writer.add_feed_forward_length(hparams.get("intermediate_size", 16384))
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if name.endswith(".A_log"):
+            data_torch = -torch.exp(data_torch)
+        elif name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+        elif name.endswith(".dt_norm_weight"):
+            name = name.rpartition(".dt_norm_weight")[0] + ".dt_norm.weight"
+        elif name.endswith(".B_norm_weight"):
+            name = name.rpartition(".B_norm_weight")[0] + ".B_norm.weight"
+        elif name.endswith(".C_norm_weight"):
+            name = name.rpartition(".C_norm_weight")[0] + ".C_norm.weight"
+        elif name.endswith(".k_weight"):
+            name = name.rpartition(".k_weight")[0] + ".k.weight"
+        elif name.endswith(".q_weight"):
+            name = name.rpartition(".q_weight")[0] + ".q.weight"
+        elif name.endswith(".conv1d.weight"):
+            data_torch = torch.squeeze(data_torch)  # remove (, 1, )
+            assert data_torch.ndim == 2
+        elif name.endswith(".pre_mixer_norm.weight"):
+            data_torch += 1.0
+        elif name.endswith(".post_mixer_norm.weight"):
+            data_torch += 1.0 / 5
+        elif name.endswith(".pre_mlp_norm.weight"):
+            data_torch += 1.0
+        elif name.endswith(".post_mlp_norm.weight"):
+            data_torch += 1.0 / (5**1.5)
+        elif name.endswith(".norm.weight"):
+            data_torch += 1.0
+
+        new_name = self.map_tensor_name(name)
+
+        return [(new_name, data_torch)]
+
+
 @ModelBase.register("CodeShellForCausalLM")
 class CodeShellModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CODESHELL
@@ -5570,7 +5903,58 @@ class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
     def set_vocab(self):
-        self._set_vocab_gpt2()
+        try:
+            self._set_vocab_gpt2()
+            return
+        except Exception:
+            pass
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        if tokpre == "kimi-k2":
+            # Build merges list using the approach similar to HunYuanMoE
+            merges = []
+            vocab = {}
+            mergeable_ranks = tokenizer.model._mergeable_ranks
+            for token, rank in mergeable_ranks.items():
+                vocab[QwenModel.token_bytes_to_string(token)] = rank
+                if len(token) == 1:
+                    continue
+                merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+                if len(merged) == 2:
+                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+
+            # Build token list
+            vocab_size = self.hparams["vocab_size"]
+            special_tokens = tokenizer.special_tokens
+            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
+            tokens: list[str] = []
+            toktypes: list[int] = []
+
+            for i in range(vocab_size):
+                if i not in reverse_vocab:
+                    tokens.append(f"[PAD{i}]")
+                    toktypes.append(gguf.TokenType.UNUSED)
+                else:
+                    token = reverse_vocab[i]
+                    tokens.append(token)
+                    if i in special_tokens.values():
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.NORMAL)
+
+            self.gguf_writer.add_tokenizer_model("gpt2")
+            self.gguf_writer.add_tokenizer_pre(tokpre)
+            self.gguf_writer.add_token_list(tokens)
+            self.gguf_writer.add_token_types(toktypes)
+            self.gguf_writer.add_token_merges(merges)
+
+            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+            special_vocab.add_to_gguf(self.gguf_writer)
+        else:
+            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
 
     def set_gguf_parameters(self):
 
@@ -6399,6 +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
@@ -146,11 +145,12 @@ pre_computed_hashes = [
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
+    {"name": "kimi-k2",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base",   "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
 ]
 
 
 def download_file_with_auth(url, token, save_path):
-    headers = {"Authorization": f"Bearer {token}"}
+    headers = {"Authorization": f"Bearer {token}"} if token else None
     response = sess.get(url, headers=headers)
     response.raise_for_status()
     os.makedirs(os.path.dirname(save_path), exist_ok=True)
@@ -231,7 +231,7 @@ for model in models:
 # generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
 
 src_ifs = ""
-for model in [*all_models, *pre_computed_hashes]:
+for model in [*pre_computed_hashes, *all_models]:
     name = model["name"]
     tokt = model["tokt"]
     chkhsh = model.get("chkhsh")
@@ -239,11 +239,6 @@ for model in [*all_models, *pre_computed_hashes]:
     if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
         continue
 
-    # Skip if the tokenizer folder does not exist or there are other download issues previously
-    if not os.path.exists(f"models/tokenizers/{name}"):
-        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
-        continue
-
     # create the tokenizer
     if chkhsh is not None:
         # if the model has a pre-computed hash, use it
@@ -253,15 +248,19 @@ for model in [*all_models, *pre_computed_hashes]:
         chkhsh = existing_models[name]
     else:
         # otherwise, compute the hash of the tokenizer
+
+        # Fail if the tokenizer folder with config does not exist or there are other download issues previously
+        if not os.path.isfile(f"models/tokenizers/{name}/tokenizer_config.json"):
+            raise OSError(f"Config for tokenizer {name} not found. The model may not exist or is not accessible with the provided token.")
+
         try:
             logger.info(f"Loading tokenizer from {f'models/tokenizers/{name}'}...")
             if name == "t5":
                 tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
             else:
                 tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-        except OSError as e:
-            logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
-            continue  # Skip to the next model if the tokenizer can't be loaded
+        except Exception as e:
+            raise OSError(f"Error loading tokenizer for model {name}.") from e
 
         chktok = tokenizer.encode(CHK_TXT)
         chkhsh = sha256(str(chktok).encode()).hexdigest()

docs/build.md

@@ -305,9 +305,8 @@ On Linux it is possible to use unified memory architecture (UMA) to share main m
 
 ## Vulkan
 
-**Windows**
-
-### w64devkit
+### For Windows Users:
+**w64devkit**
 
 Download and extract [`w64devkit`](https://github.com/skeeto/w64devkit/releases).
 
@@ -334,7 +333,7 @@ cmake -B build -DGGML_VULKAN=ON
 cmake --build build --config Release
 ```
 
-### Git Bash MINGW64
+**Git Bash MINGW64**
 
 Download and install [`Git-SCM`](https://git-scm.com/downloads/win) with the default settings
 
@@ -357,7 +356,8 @@ Now you can load the model in conversation mode using `Vulkan`
 build/bin/Release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
 ```
 
-### MSYS2
+**MSYS2**
+
 Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
 ```sh
 pacman -S git \
@@ -373,9 +373,9 @@ cmake -B build -DGGML_VULKAN=ON
 cmake --build build --config Release
 ```
 
-**With docker**:
+### For Docker users:
 
-You don't need to install Vulkan SDK. It will be installed inside the container.
+You don't need to install the Vulkan SDK. It will be installed inside the container.
 
 ```sh
 # Build the image
@@ -385,32 +385,29 @@ docker build -t llama-cpp-vulkan --target light -f .devops/vulkan.Dockerfile .
 docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
 ```
 
-**Without docker**:
+### For Linux users:
 
-Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
+First, follow the official LunarG instructions for the installation and setup of the Vulkan SDK in the [Getting Started with the Linux Tarball Vulkan SDK](https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html) guide.
 
-For example, on Ubuntu 22.04 (jammy), use the command below:
+> [!IMPORTANT]
+> After completing the first step, ensure that you have used the `source` command on the `setup_env.sh` file inside of the Vulkan SDK in your current terminal session. Otherwise, the build won't work. Additionally, if you close out of your terminal, you must perform this step again if you intend to perform a build. However, there are ways to make this persistent. Refer to the Vulkan SDK guide linked in the first step for more information about any of this.
 
+Second, after verifying that you have followed all of the SDK installation/setup steps, use this command to make sure before proceeding:
 ```bash
-wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
-wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-apt update -y
-apt-get install -y vulkan-sdk
-# To verify the installation, use the command below:
 vulkaninfo
 ```
 
-Alternatively your package manager might be able to provide the appropriate libraries.
-For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
-For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
-
-Then, build llama.cpp using the cmake command below:
-
+Then, assuming you have `cd` into your llama.cpp folder and there are no errors with running `vulkaninfo`, you can proceed to build llama.cpp using the CMake commands below:
 ```bash
 cmake -B build -DGGML_VULKAN=1
 cmake --build build --config Release
-# Test the output binary (with "-ngl 33" to offload all layers to GPU)
-./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
+```
+
+Finally, after finishing your build, you should be able to do something like this:
+```bash
+# Test the output binary
+# "-ngl 99" should offload all of the layers to GPU for most (if not all) models.
+./build/bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -ngl 99
 
 # You should see in the output, ggml_vulkan detected your GPU. For example:
 # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
@@ -557,6 +554,23 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
+## WebGPU [In Progress]
+
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
+
+In the llama.cpp directory, build with CMake:
+
+```
+cmake -B build -DGGML_WEBGPU=ON
+cmake --build build --config Release
+```
+
+### Browser Support
+
+WebGPU allows cross-platform access to the GPU from supported browsers. We utilize [Emscripten](https://emscripten.org/) to compile ggml's WebGPU backend to WebAssembly. Emscripten does not officially support WebGPU bindings yet, but Dawn currently maintains its own WebGPU bindings called emdawnwebgpu.
+
+Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/src/emdawnwebgpu/) to download or build the emdawnwebgpu package (Note that it might be safer to build the emdawbwebgpu package locally, so that it stays in sync with the version of Dawn you have installed above). When building using CMake, the path to the emdawnwebgpu port file needs to be set with the flag `EMDAWNWEBGPU_DIR`.
+
 ## IBM Z & LinuxONE
 
 To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)

examples/CMakeLists.txt

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

examples/diffusion/CMakeLists.txt

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

examples/diffusion/diffusion-cli.cpp

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

examples/embedding/embedding.cpp

@@ -107,7 +107,7 @@ int main(int argc, char ** argv) {
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     const int n_ctx_train = llama_model_n_ctx_train(model);
-    const int n_ctx = llama_n_ctx(ctx);
+    const int n_ctx       = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 

examples/parallel/parallel.cpp

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

ggml/CMakeLists.txt

@@ -181,6 +181,8 @@ option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug ou
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
+option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
+option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -270,6 +272,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-rpc.h
     include/ggml-sycl.h
     include/ggml-vulkan.h
+    include/ggml-webgpu.h
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

ggml/include/ggml-webgpu.h

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

ggml/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-cpu/CMakeLists.txt

@@ -494,9 +494,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.11.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
+        set(KLEIDIAI_ARCHIVE_MD5  "3fe9e5ab964c375c53839296eb71eaa2")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)

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

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

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

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

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

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

ggml/src/ggml-cpu/ops.cpp

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

ggml/src/ggml-cpu/vec.cpp

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

ggml/src/ggml-cuda/CMakeLists.txt

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

ggml/src/ggml-cuda/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

@@ -33,8 +33,10 @@ typedef void (* fattn_kernel_t)(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -521,7 +523,7 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
 template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
-        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne11) {
+        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11) {
     constexpr int ncols = ncols1*ncols2;
 
     const int bidx0 = blockIdx.x;
@@ -535,8 +537,8 @@ static __global__ void flash_attn_stream_k_fixup(
     const int iter_k = ne11 / FATTN_KQ_STRIDE;
     const int iter_j = (ne01 + (ncols1 - 1)) / ncols1;
 
-    const int kbc0      = (bidx0 + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
-    const int kbc0_stop = (bidx0 + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+    const int kbc0      = (bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int kbc0_stop = (bidx0 + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
 
     const bool did_not_have_any_data   = kbc0 == kbc0_stop;
     const bool wrote_beginning_of_tile = kbc0 % iter_k == 0;
@@ -545,14 +547,15 @@ static __global__ void flash_attn_stream_k_fixup(
         return;
     }
 
-    const int channel = kbc0 / (iter_k*iter_j);
-    const int jt      = (kbc0 - channel*iter_k*iter_j) / iter_k;
+    const int sequence = kbc0 / (iter_k*iter_j*(ne02/ncols2));
+    const int head = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
+    const int jt = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
 
     if (jt*ncols1 + j >= ne01) {
         return;
     }
 
-    dst += jt*ne02*(ncols1*D) + channel*(ncols2*D) + (j*ne02 + c)*D + tid;
+    dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + head*(ncols2*D) + (j*ne02 + c)*D + tid;
 
     // Load the partial result that needs a fixup:
     float dst_val = 0.0f;
@@ -571,7 +574,7 @@ static __global__ void flash_attn_stream_k_fixup(
     int bidx = bidx0 - 1;
     int kbc_stop = kbc0;
     while(true) {
-        const int kbc = bidx*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+        const int kbc = bidx*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
         if (kbc == kbc_stop) { // Did not have any data.
             bidx--;
             kbc_stop = kbc;
@@ -617,16 +620,31 @@ static __global__ void flash_attn_combine_results(
         const float2 * __restrict__ VKQ_meta,
         float * __restrict__ dst,
         const int parallel_blocks) {
-    VKQ_parts += parallel_blocks*D * gridDim.z*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.z*blockIdx.x;
-    dst       +=                 D * gridDim.z*blockIdx.x;
+    // Dimension 0: threadIdx.x
+    // Dimension 1: blockIdx.x
+    // Dimension 2: blockIdx.y
+    // Dimension 3: blockIdx.z
+    // Memory layout is permuted with [0, 2, 1, 3]
+
+    const int ne01 = gridDim.x;
+    const int ne02 = gridDim.y;
+
+    const int col      = blockIdx.x;
+    const int head     = blockIdx.y;
+    const int sequence = blockIdx.z;
+
+    const int j_dst_unrolled = (sequence*ne01 + col)*ne02 + head;
+
+    VKQ_parts += j_dst_unrolled * parallel_blocks*D;
+    VKQ_meta  += j_dst_unrolled * parallel_blocks;
+    dst       += j_dst_unrolled *                 D;
 
     const int tid = threadIdx.x;
     __builtin_assume(tid < D);
 
     extern __shared__ float2 meta[];
     for (int i = tid; i < 2*parallel_blocks; i += D) {
-        ((float *) meta)[i] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + i];
+        ((float *) meta)[i] = ((const float *)VKQ_meta) [i];
     }
 
     __syncthreads();
@@ -644,11 +662,11 @@ static __global__ void flash_attn_combine_results(
         const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
         *((uint32_t *) &KQ_max_scale) &= ftz_mask;
 
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.z*D + blockIdx.z*D + tid];
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*D + tid];
         VKQ_denominator += KQ_max_scale * meta[l].y;
     }
 
-    dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator;
+    dst[tid] = VKQ_numerator / VKQ_denominator;
 }
 
 [[noreturn]]
@@ -705,8 +723,6 @@ void launch_fattn(
 
     GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
 
-    GGML_ASSERT(Q->ne[3] == 1);
-
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t main_stream = ctx.stream();
     const int id  = ggml_cuda_get_device();
@@ -853,8 +869,8 @@ void launch_fattn(
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
         Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0,
-        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0,
+        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0, mask ? mask->ne[3] : 0,
+        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0, mask ? mask->nb[3] : 0,
         Q->nb[1], Q->nb[2], Q->nb[3],
         nb11, nb12, nb13,
         nb21, nb22, nb23,
@@ -869,11 +885,11 @@ void launch_fattn(
 
             flash_attn_stream_k_fixup<DV, ncols1, ncols2>
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]);
+                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1]);
         }
     } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(DV, 1, 1);
-        const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z);
+        const dim3 blocks_num_combine(Q->ne[1], Q->ne[2], Q->ne[3]);
         const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2);
 
         flash_attn_combine_results<DV>

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

@@ -1224,8 +1224,10 @@ static __global__ void flash_attn_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -1274,8 +1276,8 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kb_niter = FATTN_KQ_STRIDE / c::nbatch_fa; // Number of kernel iterations per assigned KQ slice.
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int       kbc      = (blockIdx.x + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
-    const int kbc_stop = (blockIdx.x + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x;
+    int       kbc      = (blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int kbc_stop = (blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
 
     // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined.
     // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup).
@@ -1285,18 +1287,19 @@ static __global__ void flash_attn_ext_f16(
     int kb0_start = kbc % iter_k;
     int kb0_stop  = min(iter_k, kb0_start + kbc_stop - kbc);
     while (kbc < kbc_stop && kb0_stop == iter_k) {
-        const int channel = kbc / (iter_k*iter_j);
-        const int jt      = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile.
+        const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
+        const int head = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
+        const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
 
-        const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
-        const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
+        const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02*(head*ncols2));
+        const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head*ncols2 / gqa_ratio));
         const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
-            (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
-        float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+            (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
+        float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head*ncols2) * (DV/2);
 
-        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head*ncols2 / gqa_ratio));
 
-        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
+        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
 
         const int kb0_start_kernel = kb0_start * kb_niter;
         const int kb0_stop_kernel  = kb0_stop  * kb_niter;
@@ -1325,18 +1328,19 @@ static __global__ void flash_attn_ext_f16(
         return;
     }
 
-    const int channel = kbc / (iter_k*iter_j);
-    const int jt      = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile.
+    const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
+    const int head = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
+    const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
 
-    const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
-    const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
+    const float2 * Q_f2    = (const float2 *) (Q + nb03*sequence + nb02*(head*ncols2));
+    const half2  * K_h2    = (const half2  *) (K + nb13*sequence + nb12*(head*ncols2 / gqa_ratio));
     const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
-        (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
-    float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+        (const half2  *) (mask + nb33*(sequence % ne33) + nb31*jt*ncols1);
+    float2       * dstk    = ((float2 *) dst) + (sequence*ne01*ne02 + head*ncols2) * (DV/2);
 
-    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head*ncols2 / gqa_ratio));
 
-    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
+    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head, n_head_log2, m0, m1) : 1.0f;
 
     const int kb0_start_kernel = kb0_start * kb_niter;
     const int kb0_stop_kernel  = kb0_stop  * kb_niter;

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

@@ -31,8 +31,10 @@ static __global__ void flash_attn_tile_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -62,15 +64,17 @@ static __global__ void flash_attn_tile_ext_f16(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const float2 * Q_f2  = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -255,6 +259,8 @@ static __global__ void flash_attn_tile_ext_f16(
         __syncthreads();
     }
 
+    float2 * dst2 = (float2 *) dst;
+
 #pragma unroll
     for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
         const int j_VKQ = j_VKQ_0 + threadIdx.y;
@@ -266,21 +272,21 @@ static __global__ void flash_attn_tile_ext_f16(
         half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
         kqsum_j = warp_reduce_sum((float)kqsum_j);
 
-#pragma unroll
-        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
-            const int i0 = i00 + 2*threadIdx.x;
+        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
 
-            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+#pragma unroll
+        for (int i00 = 0; i00 < D/2; i00 += WARP_SIZE) {
+            const int i0 = i00 + threadIdx.x;
+
+            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/WARP_SIZE];
             if (gridDim.y == 1) {
                 dst_val /= __half2half2(kqsum_j);
             }
-            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] =  __low2float(dst_val);
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = __high2float(dst_val);
+            dst2[j_dst_unrolled*(D/2) + i0] = __half22float2(dst_val);
         }
 
         if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+            dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else
@@ -290,8 +296,8 @@ static __global__ void flash_attn_tile_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
     GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
     GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
     GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);

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

@@ -31,8 +31,10 @@ static __global__ void flash_attn_tile_ext_f32(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -74,15 +76,17 @@ static __global__ void flash_attn_tile_ext_f32(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.z              + nb01*ic0);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.z / gqa_ratio));
-    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const float2 * Q_f2  = (const float2 *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
 
@@ -265,6 +269,8 @@ static __global__ void flash_attn_tile_ext_f32(
         __syncthreads();
     }
 
+    float2 * dst2 = (float2 *) dst;
+
 #pragma unroll
     for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
         const int j_VKQ = j_VKQ_0 + threadIdx.y;
@@ -276,22 +282,22 @@ static __global__ void flash_attn_tile_ext_f32(
         float kqsum_j = kqsum[j_VKQ_0/nwarps];
         kqsum_j = warp_reduce_sum(kqsum_j);
 
-#pragma unroll
-        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
-            const int i0 = i00 + 2*threadIdx.x;
+        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
 
-            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+#pragma unroll
+        for (int i00 = 0; i00 < D/2; i00 += WARP_SIZE) {
+            const int i0 = i00 + threadIdx.x;
+
+            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/WARP_SIZE];
             if (gridDim.y == 1) {
                 dst_val.x /= kqsum_j;
                 dst_val.y /= kqsum_j;
             }
-            const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = dst_val.x;
-            dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = dst_val.y;
+            dst2[j_dst_unrolled*(D/2) + i0] = dst_val;
         }
 
         if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+            dst_meta[j_dst_unrolled] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
         }
     }
 #else

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

@@ -28,8 +28,10 @@ static __global__ void flash_attn_vec_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -65,14 +67,16 @@ static __global__ void flash_attn_vec_ext_f16(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.z              + nb01*ic0;
-    K += nb12*(blockIdx.z / gqa_ratio);
-    V += nb22*(blockIdx.z / gqa_ratio);
+    Q += nb03*sequence + nb02* head              + nb01*ic0;
+    K += nb13*sequence + nb12*(head / gqa_ratio);
+    V += nb23*sequence + nb22*(head / gqa_ratio);
 
-    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
+    const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
@@ -330,12 +334,11 @@ static __global__ void flash_attn_vec_ext_f16(
         if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
+        dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + tid] = dst_val;
     }
 
     if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
+        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
@@ -344,8 +347,8 @@ static __global__ void flash_attn_vec_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
     GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
     GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
     GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);

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

@@ -28,8 +28,10 @@ static __global__ void flash_attn_vec_ext_f32(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -53,8 +55,8 @@ static __global__ void flash_attn_vec_ext_f32(
         GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
         GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
         GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
         GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
         GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
         GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -77,14 +79,16 @@ static __global__ void flash_attn_vec_ext_f32(
 
     const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb02* blockIdx.z              + nb01*ic0;
-    K += nb12*(blockIdx.z / gqa_ratio);
-    V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape
+    Q += nb03*sequence + nb02* head              + nb01*ic0;
+    K += nb13*sequence + nb12*(head / gqa_ratio);
+    V += nb23*sequence + nb22*(head / gqa_ratio);
 
-    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
+    const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
 
     static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
     constexpr int nwarps = D / WARP_SIZE;
@@ -326,12 +330,11 @@ static __global__ void flash_attn_vec_ext_f32(
         if (gridDim.y == 1) {
             dst_val /= kqsum[j_VKQ];
         }
-        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
-        dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val;
+        dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + tid] = dst_val;
     }
 
     if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
+        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
@@ -340,8 +343,8 @@ static __global__ void flash_attn_vec_ext_f32(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
     GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb32);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
     GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
     GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
     GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);

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

@@ -47,8 +47,10 @@ static __global__ void flash_attn_ext_f16(
         const int ne13,
         const int ne31,
         const int ne32,
+        const int ne33,
         const int nb31,
         const int nb32,
+        const int nb33,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -95,17 +97,19 @@ static __global__ void flash_attn_ext_f16(
     constexpr int kqs_padded = FATTN_KQ_STRIDE + 8;
     constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
 
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f   = (const float *) (Q    + nb02* blockIdx.z              + nb01*ic0);
-    const half  * K_h   = (const half  *) (K    + nb12*(blockIdx.z / gqa_ratio));
-    const half  * V_h   = (const half  *) (V    + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape
-    const half  * maskh = (const half  *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const float * Q_f   = (const float *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
+    const half  * V_h   = (const half  *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
+    const half  * maskh = (const half  *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
     const half2 * mask2 = (const half2 *)  maskh;
 
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
 
-    const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
     const half2 slope2 = make_half2(slopef, slopef);
 
@@ -400,7 +404,6 @@ static __global__ void flash_attn_ext_f16(
         if (ic0 + j_VKQ >= ne01) {
             return;
         }
-        const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y;
 
         float KQ_rowsum_j;
         if (std::is_same<KQ_acc_t, float>::value) {
@@ -409,6 +412,8 @@ static __global__ void flash_attn_ext_f16(
             KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]);
         }
 
+        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
+
 #pragma unroll
         for (int i0 = 0; i0 < D; i0 += warp_size) {
             const int i = i0 + threadIdx.x;
@@ -419,7 +424,7 @@ static __global__ void flash_attn_ext_f16(
             if (gridDim.y == 1) {
                 dst_val /= KQ_rowsum_j;
             }
-            dst[j_dst*gridDim.z*D + blockIdx.z*D + i] = dst_val;
+            dst[j_dst_unrolled*D + i] = dst_val;
         }
 
         if (gridDim.y == 1 || threadIdx.x != 0) {
@@ -433,7 +438,7 @@ static __global__ void flash_attn_ext_f16(
             dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
         }
         dst_meta_val.y = KQ_rowsum_j;
-        dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = dst_meta_val;
+        dst_meta[j_dst_unrolled] = dst_meta_val;
     }
 #else
     GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask);
@@ -442,7 +447,8 @@ static __global__ void flash_attn_ext_f16(
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
     GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
     GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
-    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32); GGML_UNUSED(ne33); GGML_UNUSED(nb31);
+    GGML_UNUSED(nb32); GGML_UNUSED(nb33); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
     GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
     GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
     GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);

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

@@ -2590,6 +2590,9 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
     // Loop over nodes in GGML graph to obtain info needed for CUDA graph
     cuda_ctx->cuda_graph->cpy_dest_ptrs.clear();
 
+    const std::string gemma3n_per_layer_proj_src0_name = "inp_per_layer_selected";
+    const std::string gemma3n_per_layer_proj_src1_name = "per_layer_proj";
+
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
 
@@ -2611,9 +2614,12 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
-            // disable CUDA graphs for batch size > 1 for now.
-            // Changes in batch size or context size can cause changes to the grid size of some kernels.
+        if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1 && (node->src[0] ? node->src[0]->name != gemma3n_per_layer_proj_src0_name : true) && (node->src[1] ? node->src[1]->name != gemma3n_per_layer_proj_src1_name : true)) {
+            // disable CUDA graphs for batch size > 1 for now while excluding the matrix-matrix addition as part of Gemma3n's `project_per_layer_input` operation
+            // by means of matching node names. See
+            // https://github.com/ggml-org/llama.cpp/blob/f9a31eea06a859e34cecb88b4d020c7f03d86cc4/src/llama-model.cpp#L10199-L10241 and
+            // https://github.com/huggingface/transformers/blob/bda75b4011239d065de84aa3e744b67ebfa7b245/src/transformers/models/gemma3n/modeling_gemma3n.py#L1773,
+            // Generally, changes in batch size or context size can cause changes to the grid size of some kernels.
             use_cuda_graph = false;
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
@@ -3226,8 +3232,9 @@ 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 +3242,9 @@ 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 +3280,6 @@ 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 +3360,7 @@ 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:
@@ -3413,12 +3410,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
-            // TODO: support broadcast
-            // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
-            //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
-            if (op->src[0]->ne[3] != 1) {
-                return false;
-            }
             if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) {
                 return false;
             }
@@ -3431,6 +3422,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
                 return true;
             }
+            if (op->src[3] && op->src[3]->ne[2] != 1) {
+                return false;
+            }
             return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) &&
                 op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
         }

ggml/src/ggml-cuda/im2col.cu

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

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

@@ -1,23 +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) {}
-
-template<>
-__device__ __forceinline__ void set_rows_1<float, half>(const float * src_f, half * dst_h) {
-    *dst_h = __float2half(*src_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, nv_bfloat16>(const float * src_f, nv_bfloat16 * dst_b) {
-    *dst_b = *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, float>(const float * src_f, float * dst_f) {
-    *dst_f = *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) {
+
+    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>
@@ -53,6 +117,9 @@ static __global__ void k_set_rows(
     const src_t* src_elem = src0_row + i00;
     dst_t* dst_elem = dst_row_ptr + i00;
     set_rows_1(src_elem, dst_elem);
+
+    GGML_UNUSED(ne10);
+    GGML_UNUSED(ne13);
 }
 
 template<typename src_t, typename dst_t>
@@ -139,7 +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

@@ -2875,12 +2875,20 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             }
 
         }
+#if GGML_SYCL_DNNL
+        // oneDNN handles strided data and does not need overhead of get_to_fp16_nc_sycl
         const int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;
         src1_f16_alloc.alloc(ne_src1);
-
         const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst);
         GGML_ASSERT(to_fp16_sycl != nullptr);
         to_fp16_sycl(src1_f16, src1_f16_alloc.get(), ne_src1, queue);
+# else
+        const int64_t ne_src1 = ggml_nelements(src1);
+        src1_f16_alloc.alloc(ne_src1);
+        const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type);
+        GGML_ASSERT(to_fp16_nc_sycl != nullptr);
+        to_fp16_nc_sycl(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, queue);
+#endif
 
         src1_f16 = src1_f16_alloc.get();
         s11      = ne10;
@@ -3522,8 +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);
@@ -3567,7 +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-sycl/set_rows.cpp

@@ -6,46 +6,49 @@ static constexpr bool is_arithmetic_v() {
     return std::is_arithmetic_v<T> || std::is_same_v<T, sycl::half> || std::is_same_v<T, sycl::ext::oneapi::bfloat16>;
 }
 }
+
 template<typename TIn, typename TOut>
 static inline std::enable_if_t<utils::is_arithmetic_v<TIn>() && utils::is_arithmetic_v<TOut>(), void>
 convert (const char* src, char* dst) {
     auto src_val = *reinterpret_cast<const TIn*>(src);
     auto dst_val = sycl::vec<TIn, 1>(src_val).template convert<TOut, sycl::rounding_mode::automatic>()[0];
-   *reinterpret_cast<TOut*>(dst) = dst_val;;
+   *reinterpret_cast<TOut*>(dst) = dst_val;
 }
 
 template<typename TIn, typename TOut>
 static void k_set_rows(
         const char * __restrict__ src0, const int64_t * __restrict__ src1, char * __restrict__ dst,
-        const int64_t ne00, const int64_t ne01, const int64_t ne11, const int64_t ne12,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t ne11, const int64_t ne12,
         const size_t nb01, const size_t nb02, const size_t nb03,
         const size_t nb10, const size_t nb11, const size_t nb12,
         const size_t nb1, const size_t nb2, const size_t nb3,
         const size_t src_type_size, const size_t dst_type_size,
-        const sycl::nd_item<3> & item_ct1) {
+        const int64_t total_elements,
+        const sycl::nd_item<1> & item_ct1) {
 
-    const int i03 = item_ct1.get_group(0);
-    const int i02 = item_ct1.get_group(1);
-    const int i01 = item_ct1.get_group(2) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1);  // Row index
-
-    if (i01 >= ne01) {
+    const int64_t i = item_ct1.get_global_linear_id();
+    if (i >= total_elements) {
         return;
     }
 
-    const int i12 = i03 % ne12;
-    const int i11 = i02 % ne11;
-    const int i10 = i01;
+    const int64_t i03 = i / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+
+    const int64_t i12 = i03 % ne12;
+    const int64_t i11 = i02 % ne11;
+    const int64_t i10 = i01;
 
     const int64_t dst_row = *(const int64_t *)((const char *)src1 + calculate_offset<3>({nb10, nb11, nb12}, {i10, i11, i12}));
 
     const char * src0_row = src0 + calculate_offset<3>({nb01, nb02, nb03}, {i01, i02, i03});
-    char * dst_row_ptr    = dst + dst_row*nb1 + i02*nb2 + i03*nb3;
+    const char * src_elem = src0_row + i00 * src_type_size;
+    char * dst_row_ptr = dst + dst_row*nb1 + i02*nb2 + i03*nb3;
+    char * dst_elem = dst_row_ptr + i00 * dst_type_size;
 
-    for (int col = item_ct1.get_local_id(0); col < ne00; col += item_ct1.get_local_range(0)) {
-        const char * src_elem = src0_row + col * src_type_size;
-        char * dst_elem       = dst_row_ptr + col * dst_type_size;
-        convert<TIn, TOut>(src_elem, dst_elem);
-    }
+    convert<TIn, TOut>(src_elem, dst_elem);
 }
 
 template<typename TIn, typename TOut>
@@ -58,33 +61,30 @@ static void set_rows_sycl(
         const size_t src_type_size, const size_t dst_type_size,
         queue_ptr stream) {
 
-    constexpr int max_threads_per_row = 64; // KEEPING 64 for now
-    const int threads_per_row     = std::min((int)ne00, max_threads_per_row);
+    const int64_t total_elements = ne00 * ne01 * ne02 * ne03;
 
-    constexpr int max_threads_per_block = 64;
-    const int rows_per_block        = std::max(1, max_threads_per_block / threads_per_row);
+    constexpr int block_size = 64;
+    const int64_t grid_size = ceil_div(total_elements, block_size);
 
-    const sycl::range<3> block_size(1, rows_per_block, threads_per_row);
-    const sycl::range<3> grid_size(ne03, ne02, (ne01 + rows_per_block - 1) / rows_per_block);
-
-        sycl_parallel_for(
-            stream,
-            sycl::nd_range<3>(grid_size * block_size, block_size),
-            [=](sycl::nd_item<3> item_ct1) {
-                k_set_rows<TIn, TOut>(
-                    src0_d, src1_d, dst_d,
-                    ne00, ne01, ne11, ne12,
-                    nb01, nb02, nb03,
-                    nb10, nb11, nb12,
-                    nb1, nb2, nb3,
-                    src_type_size, dst_type_size,
-                    item_ct1
-                );
-            }
-        );
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<1>(grid_size * block_size, block_size),
+        [=](sycl::nd_item<1> item_ct1) {
+            k_set_rows<TIn, TOut>(
+                src0_d, src1_d, dst_d,
+                ne00, ne01, ne02,
+                ne11, ne12,
+                nb01, nb02, nb03,
+                nb10, nb11, nb12,
+                nb1, nb2, nb3,
+                src_type_size, dst_type_size,
+                total_elements,
+                item_ct1
+            );
+        }
+    );
 }
 
-
 void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
     const ggml_tensor * src0 = dst->src[0];
@@ -122,7 +122,7 @@ void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
                 nb1, nb2, nb3,
                 sizeof(float), sizeof(sycl::half),
                 stream
-        );
+            );
             break;
         default:
             GGML_ABORT("Unsupported tensor type!");

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));
     };
@@ -2835,10 +2920,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         return s;
     };
 
+    bool rte = device->float_controls_rte_fp16;
 #define CREATE_BINARY(name, namemod, spec) \
     for (int s0 : {0,1}) for (int s1 : {0,1}) for (int d : {0,1}) \
         ggml_vk_create_pipeline(device, device->pipeline_ ## name ## namemod[s0][s1][d], \
-                                #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d], name ## _data[s0][s1][d], \
+                                #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d][rte], name ## _data[s0][s1][d][rte], \
                                 "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1);
 
     CREATE_BINARY(add, , {0})
@@ -2890,8 +2976,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
 #undef CREATE_UNARY
 
 #define CREATE_GLU(name)  \
-    ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);  \
-    ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);
+    if (device->float_controls_rte_fp16) {  \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);   \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16_rte", name ## _f16_rte_len, name ## _f16_rte_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);   \
+    } else {    \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);   \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);   \
+    }
 
     CREATE_GLU(geglu)
     CREATE_GLU(reglu)
@@ -2955,6 +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);
 
@@ -3267,6 +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) {
@@ -3609,6 +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) {
@@ -3629,6 +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
         }
     }
@@ -3695,10 +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;
@@ -3716,8 +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) {
@@ -4916,7 +5070,7 @@ static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
     return
         tensor->nb[0] == ggml_type_size(tensor->type) &&
         tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&
-        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+        (tensor->ne[3] == 1 || tensor->nb[3] == tensor->nb[2]*tensor->ne[2]);
 }
 
 static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src, const ggml_tensor * dst, ggml_type to) {
@@ -6803,6 +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)) {
@@ -7125,6 +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:
@@ -7991,6 +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);
@@ -9053,6 +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:
@@ -9120,6 +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:
             {
@@ -9326,6 +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);
@@ -9456,6 +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:
@@ -10007,7 +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
@@ -10037,6 +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;
@@ -10350,10 +10597,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                         // If there's not enough shared memory for row_ids and the result tile, fallback to CPU
                         return false;
                     }
-                    // Check against size of shared memory variable
-                    if (op->src[2]->ne[0] > 4096) {
-                        return false;
-                    }
                 }
                 switch (src0_type) {
                     case GGML_TYPE_F32:
@@ -10617,6 +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;
     }
@@ -11175,6 +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/copy_to_quant.comp

@@ -1,10 +1,6 @@
 #version 450
 
-#if RTE16
-#extension GL_EXT_spirv_intrinsics : enable
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif // RTE16
-
+#include "rte.comp"
 #include "types.comp"
 
 #if defined(SET_ROWS) && QUANT_K == 1

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -537,8 +537,10 @@ void process_shaders() {
     for (auto src0_f16 : {false, true}) {
     for (auto src1_f16 : {false, true}) {
     for (auto dst_f16  : {false, true}) {
-        auto name = op + get_suffix(src0_f16, src1_f16, dst_f16);
-        string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}});
+    for (auto rte      : {false, true}) {
+        auto name = op + get_suffix(src0_f16, src1_f16, dst_f16) + (rte ? "_rte" : "");
+        string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
+    }
     }
     }
     }
@@ -592,16 +594,19 @@ void process_shaders() {
     string_to_spv("sigmoid_f16",    "sigmoid.comp",     {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("sigmoid_f32",    "sigmoid.comp",     {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
-    string_to_spv("geglu_f16",      "geglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("geglu_f32",      "geglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("reglu_f16",      "reglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("reglu_f32",      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("swiglu_f16",     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("swiglu_f32",     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("geglu_erf_f16",  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("geglu_erf_f32",  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
-    string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("geglu_quick_f32","geglu_quick.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    for (auto rte : {false, true}) {
+        std::string suffix = rte ? "_rte" : "";
+        string_to_spv("geglu_f16" + suffix,      "geglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_f32" + suffix,      "geglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("reglu_f16" + suffix,      "reglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("reglu_f32" + suffix,      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_f16" + suffix,     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("swiglu_f32" + suffix,     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_erf_f16" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_erf_f32" + suffix,  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_quick_f16" + suffix,"geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("geglu_quick_f32" + suffix,"geglu_quick.comp", {{"A_TYPE", "float"},       {"D_TYPE", "float"},       {"RTE16", rte ? "1" : "0"}});
+    }
 
     string_to_spv("leaky_relu_f32", "leaky_relu.comp",  {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("silu_back_f32",  "silu_back.comp",   {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
@@ -650,6 +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"}}));
 
@@ -709,11 +716,59 @@ void write_output_files() {
             std::remove(path.c_str());
         }
     }
+
+    std::string suffixes[2] = {"_f32", "_f16"};
     for (const char *op : {"add", "sub", "mul", "div"}) {
-        fprintf(hdr, "extern unsigned char *%s_data[2][2][2];\n", op);
-        fprintf(hdr, "extern uint64_t %s_len[2][2][2];\n", op);
-        fprintf(src, "unsigned char *%s_data[2][2][2] = {{{%s_f32_f32_f32_data, %s_f32_f32_f16_data}, {%s_f32_f16_f32_data, %s_f32_f16_f16_data}}, {{%s_f16_f32_f32_data, %s_f16_f32_f16_data}, {%s_f16_f16_f32_data, %s_f16_f16_f16_data}}};\n", op, op, op, op, op, op, op, op, op);
-        fprintf(src, "uint64_t %s_len[2][2][2] = {{{%s_f32_f32_f32_len, %s_f32_f32_f16_len}, {%s_f32_f16_f32_len, %s_f32_f16_f16_len}}, {{%s_f16_f32_f32_len, %s_f16_f32_f16_len}, {%s_f16_f16_f32_len, %s_f16_f16_f16_len}}};\n", op, op, op, op, op, op, op, op, op);
+        fprintf(hdr, "extern unsigned char *%s_data[2][2][2][2];\n", op);
+        fprintf(hdr, "extern uint64_t %s_len[2][2][2][2];\n", op);
+        std::string data = "unsigned char *" + std::string(op) + "_data[2][2][2][2] = ";
+        std::string len = "uint64_t " + std::string(op) + "_len[2][2][2][2] = ";
+        for (uint32_t t0 = 0; t0 < 2; ++t0) {
+            if (t0 == 0) {
+                data += "{";
+                len += "{";
+            }
+            for (uint32_t t1 = 0; t1 < 2; ++t1) {
+                if (t1 == 0) {
+                    data += "{";
+                    len += "{";
+                }
+                for (uint32_t t2 = 0; t2 < 2; ++t2) {
+                    if (t2 == 0) {
+                        data += "{";
+                        len += "{";
+                    }
+                    for (uint32_t rte = 0; rte < 2; ++rte) {
+                        if (rte == 0) {
+                            data += "{";
+                            len += "{";
+                        }
+                        data += op + suffixes[t0] + suffixes[t1] + suffixes[t2] + ((rte != 0) ? "_rte" : "");
+                        len  += op + suffixes[t0] + suffixes[t1] + suffixes[t2] + ((rte != 0) ? "_rte" : "");
+                        data += "_data,";
+                        len  += "_len,";
+                        if (rte == 1) {
+                            data += "}, ";
+                            len += "}, ";
+                        }
+                    }
+                    if (t2 == 1) {
+                        data += "}, ";
+                        len += "}, ";
+                    }
+                }
+                if (t1 == 1) {
+                    data += "}, ";
+                    len += "}, ";
+                }
+            }
+            if (t0 == 1) {
+                data += "};\n";
+                len += "};\n";
+            }
+        }
+        fputs(data.c_str(), src);
+        fputs(len.c_str(), src);
     }
     fclose(hdr);
     fclose(src);

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
 
 
@@ -317,6 +323,7 @@ class MODEL_ARCH(IntEnum):
     PHI3             = auto()
     PHIMOE           = auto()
     PLAMO            = auto()
+    PLAMO2           = auto()
     CODESHELL        = auto()
     ORION            = auto()
     INTERNLM2        = auto()
@@ -353,6 +360,7 @@ class MODEL_ARCH(IntEnum):
     JAIS             = auto()
     NEMOTRON         = auto()
     EXAONE           = auto()
+    EXAONE4          = auto()
     GRANITE          = auto()
     GRANITE_MOE      = auto()
     GRANITE_HYBRID   = auto()
@@ -363,9 +371,11 @@ class MODEL_ARCH(IntEnum):
     DOTS1            = auto()
     ARCEE            = auto()
     ERNIE4_5         = auto()
+    ERNIE4_5_MOE     = auto()
     HUNYUAN_MOE      = auto()
     SMOLLM3          = auto()
     LFM2             = auto()
+    DREAM            = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -631,6 +641,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PHI3:             "phi3",
     MODEL_ARCH.PHIMOE:           "phimoe",
     MODEL_ARCH.PLAMO:            "plamo",
+    MODEL_ARCH.PLAMO2:           "plamo2",
     MODEL_ARCH.CODESHELL:        "codeshell",
     MODEL_ARCH.ORION:            "orion",
     MODEL_ARCH.INTERNLM2:        "internlm2",
@@ -667,6 +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",
@@ -677,10 +689,12 @@ 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",
     MODEL_ARCH.LFM2:             "lfm2",
+    MODEL_ARCH.DREAM:            "dream",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1287,6 +1301,21 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.DREAM: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        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_ARCH.QWEN2VL: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1369,6 +1398,36 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.PLAMO2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_POST_NORM,
+        MODEL_TENSOR.SSM_IN,
+        MODEL_TENSOR.SSM_CONV1D,
+        MODEL_TENSOR.SSM_X,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_D,
+        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.SSM_DT_NORM,
+        MODEL_TENSOR.SSM_B_NORM,
+        MODEL_TENSOR.SSM_C_NORM,
+    ],
     MODEL_ARCH.GPT2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.POS_EMBD,
@@ -1973,6 +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,
@@ -2124,6 +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/scripts/gguf_dump.py

@@ -234,6 +234,8 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
     markdown_content += '## Key Value Metadata Store\n\n'
     markdown_content += f'There are {len(reader.fields)} key-value pairs in this file\n'
     markdown_content += '\n'
+    total_model_bytes = 0
+    total_model_elements = 0
 
     kv_dump_table: list[dict[str, str | int]] = []
     for n, field in enumerate(reader.fields.values(), 1):
@@ -377,6 +379,8 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
             tensors = tensor_groups[group]
             group_elements = sum(tensor.n_elements for tensor in tensors)
             group_percentage = group_elements / total_elements * 100
+            total_group_bytes = 0
+            total_group_elements = 0
             markdown_content += f"### <a name=\"{group.replace('.', '_')}\">{translate_tensor_name(group)} Tensor Group : {element_count_rounded_notation(group_elements)} Elements</a>\n\n"
 
             # Precalculate column sizing for visual consistency
@@ -397,7 +401,13 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
                 element_count_est = f"({element_count_rounded_notation(tensor.n_elements):>{prettify_element_est_count_size}})"
                 element_count_string = f"{element_count_est} {tensor.n_elements:>{prettify_element_count_size}}"
                 type_name_string = f"{tensor.tensor_type.name}"
-                tensor_dump_table.append({"t_id":tensor_name_to_key[tensor.name], "layer_name":tensor.name, "human_layer_name":human_friendly_name, "element_count":element_count_string, "pretty_dimension":pretty_dimension, "tensor_type":type_name_string})
+                if tensor.n_elements > 0:
+                    bpw = (tensor.n_bytes * 8) / tensor.n_elements
+                else:
+                    bpw = float('nan')
+                tensor_dump_table.append({"t_id":tensor_name_to_key[tensor.name], "layer_name":tensor.name, "human_layer_name":human_friendly_name, "element_count":element_count_string, "pretty_dimension":pretty_dimension, "tensor_type":type_name_string, "bpw": f"{bpw:.4f}"})
+                total_group_bytes += tensor.n_bytes
+                total_group_elements += tensor.n_elements
 
             tensor_dump_table_header_map = [
                 {'key_name':'t_id',             'header_name':'T_ID',                             'align':'right'},
@@ -406,6 +416,7 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
                 {'key_name':'element_count',    'header_name':'Elements',                         'align':'left'},
                 {'key_name':'pretty_dimension', 'header_name':'Shape',                            'align':'left'},
                 {'key_name':'tensor_type',      'header_name':'Type',                             'align':'left'},
+                {'key_name':'bpw',              'header_name':'BPW',                              'align':'right'},
             ]
 
             markdown_content += markdown_table_with_alignment_support(tensor_dump_table_header_map, tensor_dump_table)
@@ -413,8 +424,20 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
             markdown_content += "\n"
             markdown_content += f"- Total elements in {group}: ({element_count_rounded_notation(group_elements):>4}) {group_elements}\n"
             markdown_content += f"- Percentage of total elements: {group_percentage:.2f}%\n"
+            if total_group_elements > 0:
+                total_group_bpw = (total_group_bytes * 8) / total_group_elements
+                markdown_content += f"- Bits per Weight (BPW) for {group}: {total_group_bpw:.4f} bits\n"
+            else:
+                markdown_content += f"- Bits per Weight (BPW) for {group}: undefined (no elements)\n"
             markdown_content += "\n\n"
+            total_model_bytes += total_group_bytes
+            total_model_elements += total_group_elements
 
+    if total_model_elements > 0:
+        total_model_bpw = (total_model_bytes * 8) / total_model_elements
+        markdown_content += f"Total BPW for {os.path.basename(args.model)}: {total_model_bpw:.4f} bits"
+    else:
+        markdown_content += f"Total BPW for {os.path.basename(args.model)}: undefined (no elements)"
     print(markdown_content)  # noqa: NP100
 
 

gguf-py/gguf/tensor_mapping.py

@@ -13,7 +13,7 @@ class TensorNameMap:
             "transformer.wte",                           # gpt2 gpt-j mpt refact qwen dbrx jais exaone
             "transformer.word_embeddings",               # falcon
             "word_embeddings",                           # bloom
-            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414 granite-hybrid
+            "model.embed_tokens",                        # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414 plamo2 granite-hybrid
             "tok_embeddings",                            # llama-pth
             "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
@@ -63,7 +63,7 @@ class TensorNameMap:
         # Output
         MODEL_TENSOR.OUTPUT: (
             "embed_out",                 # gptneox
-            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2 phimoe
+            "lm_head",                   # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe olmo2 phimoe plamo2
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
@@ -77,7 +77,7 @@ class TensorNameMap:
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
             "transformer.ln_f",                        # gpt2 gpt-j falcon jais exaone
-            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo2 phimoe
+            "model.norm",                              # llama-hf baichuan internlm2 olmoe olmo2 phimoe plamo2
             "norm",                                    # llama-pth
             "transformer.norm_f",                      # mpt dbrx
             "ln_f",                                    # refact bloom qwen gpt2
@@ -126,6 +126,7 @@ class TensorNameMap:
             "h.{bid}.ln_1",                                         # gpt2
             "transformer.h.{bid}.ln",                               # phi2
             "model.layers.layers.{bid}.norm",                       # plamo
+            "model.layers.layers.{bid}.pre_mixer_norm",             # plamo2
             "model.layers.{bid}.attention_norm",                    # internlm2
             "model.layers.{bid}.norm",                              # mamba-qbert
             "backbone.layers.{bid}.norm",                           # mamba
@@ -163,6 +164,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
             "encoder.layers.{bid}.mixer.Wqkv",                                     # jina
             "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
+            "model.layers.layers.{bid}.mixer.qkv_proj",                            # plamo2
             "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
             "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
             "transformer_encoder.{bid}.qkv",                                       # neobert
@@ -233,6 +235,7 @@ class TensorNameMap:
             "h.{bid}.attn.c_proj",                                          # gpt2
             "transformer.h.{bid}.mixer.out_proj",                           # phi2
             "model.layers.layers.{bid}.self_attn.o_proj",                   # plamo
+            "model.layers.layers.{bid}.mixer.o_proj",                       # plamo2
             "model.layers.{bid}.attention.wo",                              # internlm2
             "encoder.layers.{bid}.attn.out_proj",                           # nomic-bert
             "encoder.layers.{bid}.mixer.out_proj",                          # jina
@@ -255,8 +258,9 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.ATTN_POST_NORM: (
-            "model.layers.{bid}.post_attention_layernorm",     # gemma2 olmo2    # ge
-            "model.layers.{bid}.post_self_attn_layernorm",     # glm-4-0414
+            "model.layers.{bid}.post_attention_layernorm",       # gemma2 olmo2    # ge
+            "model.layers.{bid}.post_self_attn_layernorm",       # glm-4-0414
+            "model.layers.layers.{bid}.post_mixer_norm.weight",  # plamo2
         ),
 
         # Rotary embeddings
@@ -286,6 +290,7 @@ class TensorNameMap:
             "model.layers.{bid}.pre_moe_layernorm",                          # mini-jamba
             "model.layers.{bid}.post_attention_layernorm",                   # llama4
             "transformer_encoder.{bid}.ffn_norm",                            # neobert
+            "model.layers.layers.{bid}.pre_mlp_norm",                        # plamo2
         ),
 
         # Post feed-forward norm
@@ -298,6 +303,7 @@ class TensorNameMap:
         MODEL_TENSOR.FFN_POST_NORM: (
             "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2
             "model.layers.{bid}.post_mlp_layernorm", # glm-4-0414
+            "model.layers.layers.{bid}.post_mlp_norm.weight", # plamo2
             "model.layers.{bid}.feed_forward.up_proj",
         ),
 
@@ -318,7 +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
@@ -342,6 +349,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.fc1",                             # phi2
             "model.layers.{bid}.mlp.gate_up_proj",                    # phi3 glm-4-0414
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
+            "model.layers.layers.{bid}.mlp.gate_up_proj",             # plamo2
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "encoder.layers.{bid}.mlp.fc1",                           # nomic-bert-moe
@@ -357,13 +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: (
@@ -396,12 +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: (
@@ -443,14 +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: (
@@ -469,6 +477,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
+            "model.layers.layers.{bid}.mixer.q",                              # plamo2
         ),
 
         MODEL_TENSOR.ATTN_K_NORM: (
@@ -479,6 +488,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm
+            "model.layers.layers.{bid}.mixer.k",                              # plamo2
         ),
 
         MODEL_TENSOR.ROPE_FREQS: (
@@ -559,27 +569,31 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.SSM_IN: (
-            "model.layers.{bid}.in_proj",           # mamba-hf
-            "backbone.layers.{bid}.mixer.in_proj",  # mamba
-            "model.layers.{bid}.mamba.in_proj",     # jamba falcon-h1 granite-hybrid
+            "model.layers.{bid}.in_proj",               # mamba-hf
+            "backbone.layers.{bid}.mixer.in_proj",      # mamba
+            "model.layers.{bid}.mamba.in_proj",         # jamba falcon-h1 granite-hybrid
+            "model.layers.layers.{bid}.mixer.in_proj",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_CONV1D: (
-            "model.layers.{bid}.conv1d",           # mamba-hf
-            "backbone.layers.{bid}.mixer.conv1d",  # mamba
-            "model.layers.{bid}.mamba.conv1d",     # jamba falcon-h1 granite-hybrid
+            "model.layers.{bid}.conv1d",               # mamba-hf
+            "backbone.layers.{bid}.mixer.conv1d",      # mamba
+            "model.layers.{bid}.mamba.conv1d",         # jamba falcon-h1 granite-hybrid
+            "model.layers.layers.{bid}.mixer.conv1d",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_X: (
-            "model.layers.{bid}.x_proj",           # mamba-hf
-            "backbone.layers.{bid}.mixer.x_proj",  # mamba
-            "model.layers.{bid}.mamba.x_proj",     # jamba
+            "model.layers.{bid}.x_proj",                  # mamba-hf
+            "backbone.layers.{bid}.mixer.x_proj",         # mamba
+            "model.layers.{bid}.mamba.x_proj",            # jamba
+            "model.layers.layers.{bid}.mixer.bcdt_proj",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_DT: (
-            "model.layers.{bid}.dt_proj",           # mamba-hf
-            "backbone.layers.{bid}.mixer.dt_proj",  # mamba
-            "model.layers.{bid}.mamba.dt_proj",     # jamba falcon-h1 granite-hybrid
+            "model.layers.{bid}.dt_proj",               # mamba-hf
+            "backbone.layers.{bid}.mixer.dt_proj",      # mamba
+            "model.layers.{bid}.mamba.dt_proj",         # jamba falcon-h1 granite-hybrid
+            "model.layers.layers.{bid}.mixer.dt_proj",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_DT_NORM: (
@@ -587,25 +601,33 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.SSM_A: (
-            "model.layers.{bid}.A_log",           # mamba-hf
-            "backbone.layers.{bid}.mixer.A_log",  # mamba
-            "model.layers.{bid}.mamba.A_log",     # jamba falcon-h1 granite-hybrid
+            "model.layers.{bid}.A_log",               # mamba-hf
+            "backbone.layers.{bid}.mixer.A_log",      # mamba
+            "model.layers.{bid}.mamba.A_log",         # jamba falcon-h1 granite-hybrid
+            "model.layers.layers.{bid}.mixer.A_log",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_B_NORM: (
-            "model.layers.{bid}.mamba.b_layernorm",  # jamba
-            "model.layers.{bid}.mamba.B_layernorm",  # mini-jamba
+            "model.layers.{bid}.mamba.b_layernorm",           # jamba
+            "model.layers.{bid}.mamba.B_layernorm",           # mini-jamba
+            "model.layers.layers.{bid}.mixer.B_norm.weight",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_C_NORM: (
-            "model.layers.{bid}.mamba.c_layernorm",  # jamba
-            "model.layers.{bid}.mamba.C_layernorm",  # mini-jamba
+            "model.layers.{bid}.mamba.c_layernorm",           # jamba
+            "model.layers.{bid}.mamba.C_layernorm",           # mini-jamba
+            "model.layers.layers.{bid}.mixer.C_norm.weight",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_D: (
-            "model.layers.{bid}.D",           # mamba-hf
-            "backbone.layers.{bid}.mixer.D",  # mamba
-            "model.layers.{bid}.mamba.D",     # jamba falcon-h1 granite-hybrid
+            "model.layers.{bid}.D",               # mamba-hf
+            "backbone.layers.{bid}.mixer.D",      # mamba
+            "model.layers.{bid}.mamba.D",         # jamba falcon-h1 granite-hybrid
+            "model.layers.layers.{bid}.mixer.D",  # plamo2
+        ),
+
+        MODEL_TENSOR.SSM_DT_NORM: (
+            "model.layers.layers.{bid}.mixer.dt_norm.weight",  # plamo2
         ),
 
         MODEL_TENSOR.SSM_NORM: (
@@ -614,9 +636,10 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.SSM_OUT: (
-            "model.layers.{bid}.out_proj",           # mamba-hf
-            "backbone.layers.{bid}.mixer.out_proj",  # mamba
-            "model.layers.{bid}.mamba.out_proj",     # jamba falcon-h1 granite-hybrid
+            "model.layers.{bid}.out_proj",               # mamba-hf
+            "backbone.layers.{bid}.mixer.out_proj",      # mamba
+            "model.layers.{bid}.mamba.out_proj",         # jamba falcon-h1 granite-hybrid
+            "model.layers.layers.{bid}.mixer.out_proj",  # plamo2
         ),
 
         MODEL_TENSOR.TIME_MIX_W0: (

include/llama.h

@@ -71,12 +71,13 @@ extern "C" {
     typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
-        LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
-        LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
-        LLAMA_VOCAB_TYPE_UGM  = 4, // T5 tokenizer based on Unigram
-        LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization
+        LLAMA_VOCAB_TYPE_NONE   = 0, // For models without vocab
+        LLAMA_VOCAB_TYPE_SPM    = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
+        LLAMA_VOCAB_TYPE_BPE    = 2, // GPT-2 tokenizer based on byte-level BPE
+        LLAMA_VOCAB_TYPE_WPM    = 3, // BERT tokenizer based on WordPiece
+        LLAMA_VOCAB_TYPE_UGM    = 4, // T5 tokenizer based on Unigram
+        LLAMA_VOCAB_TYPE_RWKV   = 5, // RWKV tokenizer based on greedy tokenization
+        LLAMA_VOCAB_TYPE_PLAMO2 = 6, // PLaMo-2 tokenizer based on Aho-Corasick with dynamic programming
     };
 
     enum llama_rope_type {
@@ -334,6 +335,9 @@ extern "C" {
         bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
                           // NOTE: setting to false when n_seq_max > 1 can cause bad performance in some cases
                           //       ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573
+        bool kv_unified;  // use a unified buffer across the input sequences when computing the attention
+                          // try to disable when n_seq_max > 1 for improved performance when the sequences do not share a large prefix
+                          // ref: https://github.com/ggml-org/llama.cpp/pull/14363
     };
 
     // model quantization parameters
@@ -724,7 +728,7 @@ extern "C" {
     //   - lazily on next llama_decode()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    DEPRECATED(void llama_kv_self_seq_div(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
@@ -1004,6 +1008,7 @@ extern "C" {
     LLAMA_API llama_token llama_vocab_sep(const struct llama_vocab * vocab); // sentence separator
     LLAMA_API llama_token llama_vocab_nl (const struct llama_vocab * vocab); // next-line
     LLAMA_API llama_token llama_vocab_pad(const struct llama_vocab * vocab); // padding
+    LLAMA_API llama_token llama_vocab_mask(const struct llama_vocab * vocab); // mask
 
     LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab);
     LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab);
@@ -1389,6 +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 {

models/templates/moonshotai-Kimi-K2.jinja

@@ -0,0 +1,43 @@
+{%- if tools -%}
+  <|im_system|>tool_declare<|im_middle|>{{ tools | tojson }}<|im_end|>
+{%- endif -%}
+{%- for message in messages -%}
+  {%- if loop.first and messages[0]['role'] != 'system' -%}
+    <|im_system|>system<|im_middle|>You are a helpful assistant<|im_end|>
+  {%- endif -%}
+  {%- if message['role'] == 'system' -%}
+    <|im_system|>system<|im_middle|>
+  {%- elif message['role'] == 'user' -%}
+    <|im_user|>user<|im_middle|>
+  {%- elif message['role'] == 'assistant' -%}
+    <|im_assistant|>assistant<|im_middle|>
+  {%- elif message['role'] == 'tool' -%}
+    <|im_system|>tool<|im_middle|>
+  {%- endif -%}
+  {%- if message['role'] == 'assistant' and message.get('tool_calls') -%}
+    {%- if message['content'] -%}{{ message['content'] }}{%- endif -%}
+    <|tool_calls_section_begin|>
+    {%- for tool_call in message['tool_calls'] -%}
+      {%- set func_name = tool_call['function']['name'] -%}
+      {%- set formatted_id = 'functions.' + func_name + ':' + loop.index0|string -%}
+      <|tool_call_begin|>{{ formatted_id }}<|tool_call_argument_begin|>{{ tool_call['function']['arguments'] | tojson}}<|tool_call_end|>
+    {%- endfor -%}
+    <|tool_calls_section_end|>
+  {%- elif message['role'] == 'tool' -%}
+    ## Return of {{ message.tool_call_id }}\n{{ message['content'] }}
+  {%- elif message['content'] is string -%}
+    {{ message['content'] }}
+  {%- elif message['content'] is not none -%}
+    {% for content in message['content'] -%}
+      {% if content['type'] == 'image' or 'image' in content or 'image_url' in content -%}
+        <|media_start|>image<|media_content|><|media_pad|><|media_end|>
+      {% else -%}
+        {{ content['text'] }}
+      {%- endif -%}
+    {%- endfor -%}
+  {%- endif -%}
+  <|im_end|>
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  <|im_assistant|>assistant<|im_middle|>
+{%- endif -%}

requirements/requirements-all.txt

@@ -3,6 +3,7 @@
 -r ../tools/server/tests/requirements.txt
 
 -r ./requirements-compare-llama-bench.txt
+-r ./requirements-server-bench.txt
 -r ./requirements-pydantic.txt
 -r ./requirements-test-tokenizer-random.txt
 

requirements/requirements-server-bench.txt

@@ -0,0 +1,5 @@
+datasets~=3.2.0
+matplotlib~=3.10.0
+numpy~=1.26.4
+requests~=2.32.3
+tqdm~=4.67.1

scripts/server-bench.py

@@ -0,0 +1,265 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+import os
+import random
+import subprocess
+from time import sleep, time
+from typing import Optional, Union
+
+import datasets
+import logging
+import matplotlib.pyplot as plt
+import numpy as np
+import requests
+from tqdm.contrib.concurrent import thread_map
+
+
+logging.basicConfig(level=logging.INFO, format='%(message)s')
+logger = logging.getLogger("server-bench")
+
+
+def get_prompts_text(dataset_name: str, n_prompts: int) -> Optional[list[str]]:
+    ret = []
+    if dataset_name.lower() == "mmlu":
+        logger.info("Loading MMLU dataset...")
+        ret = datasets.load_dataset("cais/mmlu", "all")["test"]["question"]  # type: ignore
+    else:
+        return None
+    if n_prompts >= 0:
+        ret = ret[:n_prompts]
+    return ret
+
+
+def get_prompt_lengths_rng(n_prompts: int, prompt_length_min: int, prompt_length_max: int) -> list[int]:
+    assert n_prompts >= 0
+    ret: list[int] = []
+    for i in range(n_prompts):
+        random.seed(13 * i + 0)
+        ret.append(random.randint(prompt_length_min, prompt_length_max))
+    return ret
+
+
+def get_prompts_rng(prompt_lengths: list[int]) -> list[list[int]]:
+    return [[random.randint(100, 10000) for _ in range(pl)] for pl in prompt_lengths]
+
+
+def get_server(path_server: str, path_log: Optional[str]) -> dict:
+    logger.info("Starting the llama.cpp server...")
+    hostname: str = os.environ.get("LLAMA_ARG_HOST", "127.0.0.1")
+    port: str = os.environ.get("LLAMA_ARG_PORT", "8080")
+    address: str = f"http://{hostname}:{port}"
+
+    fout = open(path_log, "w") if path_log is not None else subprocess.DEVNULL
+    process = subprocess.Popen([path_server], stdout=fout, stderr=subprocess.STDOUT)
+
+    n_failures: int = 0
+    while True:
+        try:
+            sleep(1.0)
+            exit_code = process.poll()
+            if exit_code is not None:
+                raise RuntimeError(f"llama.cpp server exited unexpectedly with exit code {exit_code}, see {path_log}")
+            response = requests.get(f"{address}/health")
+            if response.status_code == 200:
+                break
+        except requests.ConnectionError:
+            n_failures += 1
+            if n_failures >= 10:
+                raise RuntimeError("llama.cpp server is not healthy after 10 seconds")
+
+    return {"process": process, "address": address, "fout": fout}
+
+
+def get_prompt_length(data: dict) -> int:
+    session = data["session"]
+    server_address: str = data["server_address"]
+
+    response = session.post(
+        f"{server_address}/apply-template",
+        json={"messages": [{"role": "user", "content": data["prompt"], "stream": True}]}
+    )
+    if response.status_code != 200:
+        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+    prompt: str = json.loads(response.text)["prompt"]
+    response = session.post(
+        f"{server_address}/tokenize",
+        json={"content": prompt, "add_special": True}
+    )
+    if response.status_code != 200:
+        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+    tokens: list[str] = json.loads(response.text)["tokens"]
+    return len(tokens)
+
+
+def send_prompt(data: dict) -> tuple[float, list[float]]:
+    session = data["session"]
+    server_address: str = data["server_address"]
+
+    t_submit = time()
+    if data["synthetic_prompt"]:
+        json_data: dict = {
+            "prompt": data["prompt"], "ignore_eos": True, "cache_prompt": False,
+            "seed": data["seed"], "n_predict": data["n_predict"], "stream": True}
+        response = session.post(f"{server_address}/completion", json=json_data, stream=True)
+    else:
+        response = session.post(
+            f"{server_address}/apply-template",
+            json={"messages": [{"role": "user", "content": data["prompt"], "stream": True}]}
+        )
+        if response.status_code != 200:
+            raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+        prompt: str = json.loads(response.text)["prompt"]
+
+        json_data: dict = {"prompt": prompt, "seed": data["seed"], "n_predict": data["n_predict"], "stream": True}
+        response = session.post(f"{server_address}/completion", json=json_data, stream=True)
+
+    token_arrival_times: list[float] = []
+    for line in response.iter_lines(decode_unicode=False):
+        if not line.startswith(b"data: "):
+            continue
+        token_arrival_times.append(time())
+    token_arrival_times = token_arrival_times[:-1]
+
+    if response.status_code != 200:
+        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+
+    return (t_submit, token_arrival_times)
+
+
+def benchmark(path_server: str, path_log: Optional[str], prompt_source: str, n_prompts: int, n_predict: int, n_predict_min: int):
+    if os.environ.get("LLAMA_ARG_N_PARALLEL") is None:
+        logger.info("LLAMA_ARG_N_PARALLEL not explicitly set, using 32")
+        os.environ["LLAMA_ARG_N_PARALLEL"] = "32"
+    if os.environ.get("LLAMA_ARG_N_GPU_LAYERS") is None:
+        logger.info("LLAMA_ARG_N_GPU_LAYERS not explicitly set, using 999")
+        os.environ["LLAMA_ARG_N_GPU_LAYERS"] = "999"
+    if os.environ.get("LLAMA_ARG_FLASH_ATTN") is None:
+        logger.info("LLAMA_ARG_FLASH_ATTN not explicitly set, using 'true'")
+        os.environ["LLAMA_ARG_FLASH_ATTN"] = "true"
+
+    parallel: int = int(os.environ.get("LLAMA_ARG_N_PARALLEL", 1))
+    prompts: Union[None, list[str], list[list[int]]] = get_prompts_text(prompt_source, n_prompts)
+    synthetic_prompts: bool = prompts is None
+    prompt_n = []
+
+    if synthetic_prompts:
+        prompt_source_split: list[str] = prompt_source.split("-")
+        assert len(prompt_source_split) == 3
+        assert prompt_source_split[0].lower() == "rng"
+        prompt_length_min: int = int(prompt_source_split[1])
+        prompt_length_max: int = int(prompt_source_split[2])
+        logger.info("Generating random prompts...")
+        prompt_n = get_prompt_lengths_rng(n_prompts, prompt_length_min, prompt_length_max)
+        prompts = get_prompts_rng(prompt_n)
+    else:
+        n_predict_min = n_predict
+
+    if os.environ.get("LLAMA_ARG_CTX_SIZE") is None:
+        context_per_slot: int = int(1.05 * (n_predict + (np.max(prompt_n) if synthetic_prompts else 2048)))
+        context_total: int = context_per_slot * parallel
+        os.environ["LLAMA_ARG_CTX_SIZE"] = str(context_total)
+        logger.info(f"LLAMA_ARG_CTX_SIZE not explicitly set, using {context_total} ({context_per_slot} per slot).")
+
+    server: Optional[dict] = None
+    session = None
+    try:
+        server = get_server(path_server, path_log)
+        server_address: str = server["address"]
+
+        adapter = requests.adapters.HTTPAdapter(pool_connections=parallel, pool_maxsize=parallel)  # type: ignore
+        session = requests.Session()
+        session.mount("http://", adapter)
+        session.mount("https://", adapter)
+
+        data: list[dict] = []
+
+        for i, p in enumerate(prompts):
+            random.seed(13 * i + 1)
+            data.append({
+                "session": session, "server_address": server_address, "prompt": p, "synthetic_prompt": synthetic_prompts,
+                "n_predict": random.randint(n_predict_min, n_predict), "seed": 13 * i + 2})
+
+        if not synthetic_prompts:
+            logger.info("Getting the prompt lengths...")
+            prompt_n = [get_prompt_length(d) for d in data]
+
+        logger.info("Starting the benchmark...\n")
+        t0 = time()
+        results: list[tuple[float, list[float]]] = thread_map(send_prompt, data, max_workers=parallel, chunksize=1)
+    finally:
+        if server is not None:
+            server["process"].terminate()
+            server["process"].wait()
+        if session is not None:
+            session.close()
+
+    prompt_t = []
+    token_t = []
+    depth_sum: int = 0
+    for pn, (t_submit, tat) in zip(prompt_n, results):
+        prompt_t.append(tat[0] - t_submit)
+        token_t += tat
+        n_tokens: int = len(tat)
+        depth_sum += n_tokens * pn
+        depth_sum += n_tokens * (n_tokens + 1) // 2
+    assert len(token_t) > 0
+    prompt_n = np.array(prompt_n, dtype=np.int64)
+    prompt_t = np.array(prompt_t, dtype=np.float64)
+    token_t = np.array(token_t, dtype=np.float64)
+
+    token_t -= t0
+    token_t_last = np.max(token_t)
+
+    logger.info("")
+    logger.info(f"Benchmark duration:                {token_t_last:.2f} s")
+    logger.info(f"Request throughput:                {n_prompts / token_t_last:.2f} requests/s = {n_prompts / (token_t_last/60):.2f} requests/min")
+    logger.info(f"Total prompt length:               {np.sum(prompt_n)} tokens")
+    logger.info(f"Average prompt length:             {np.mean(prompt_n):.2f} tokens")
+    logger.info(f"Average prompt latency:            {1e3 * np.mean(prompt_t):.2f} ms")
+    logger.info(f"Average prompt speed:              {np.sum(prompt_n) / np.sum(prompt_t):.2f} tokens/s")
+    logger.info(f"Total generated tokens:            {token_t.shape[0]}")
+    logger.info(f"Average generation depth:          {depth_sum / token_t.shape[0]:.2f} tokens")
+    logger.info(f"Average total generation speed:    {token_t.shape[0] / token_t_last:.2f} tokens/s")
+    logger.info(f"Average generation speed per slot: {token_t.shape[0] / (parallel * token_t_last):.2f} tokens/s / slot")
+    logger.info("")
+    logger.info(
+        "The above numbers are the speeds as observed by the Python script and may differ from the performance reported by the server, "
+        "particularly when the server is fast vs. the network or Python script (e.g. when serving a very small model).")
+
+    plt.figure()
+    plt.scatter(prompt_n, 1e3 * prompt_t, s=10.0, marker=".", alpha=0.25)
+    plt.xlim(0, 1.05e0 * np.max(prompt_n))
+    plt.ylim(0, 1.05e3 * np.max(prompt_t))
+    plt.xlabel("Prompt length [tokens]")
+    plt.ylabel("Time to first token [ms]")
+    plt.savefig("prompt_time.png", dpi=240)
+
+    bin_max = np.ceil(token_t_last) + 1
+    plt.figure()
+    plt.hist(token_t, np.arange(0, bin_max))
+    plt.xlim(0, bin_max + 1)
+    plt.xlabel("Time [s]")
+    plt.ylabel("Num. tokens generated per second")
+    plt.savefig("gen_rate.png", dpi=240)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Tool for benchmarking the throughput of the llama.cpp HTTP server. "
+        "Results are printed to console and visualized as plots (saved to current working directory). "
+        "To pass arguments such as the model path to the server, set the corresponding environment variables (see llama-server --help).")
+    parser.add_argument("--path_server", type=str, default="llama-server", help="Path to the llama.cpp server binary")
+    parser.add_argument("--path_log", type=str, default="server-bench.log", help="Path to the model to use for the benchmark")
+    parser.add_argument(
+        "--prompt_source", type=str, default="rng-1024-2048",
+        help="How to get the prompts for the benchmark, either 'mmlu' for MMLU questions or "
+        "rng-MIN-MAX for synthetic prompts with random lengths in the interval [MIN, MAX]")
+    parser.add_argument("--n_prompts", type=int, default=100, help="Number of prompts to evaluate")
+    parser.add_argument("--n_predict", type=int, default=2048, help="Max. number of tokens to predict per prompt")
+    parser.add_argument(
+        "--n_predict_min", type=int, default=1024,
+        help="Min. number of tokens to predict per prompt (supported for synthetic prompts only)")
+    args = parser.parse_args()
+    benchmark(**vars(args))

scripts/sync-ggml.last

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

src/llama-arch.cpp

@@ -34,6 +34,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_PHI3,             "phi3"             },
     { LLM_ARCH_PHIMOE,           "phimoe"           },
     { LLM_ARCH_PLAMO,            "plamo"            },
+    { LLM_ARCH_PLAMO2,           "plamo2"           },
     { LLM_ARCH_CODESHELL,        "codeshell"        },
     { LLM_ARCH_ORION,            "orion"            },
     { LLM_ARCH_INTERNLM2,        "internlm2"        },
@@ -67,6 +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"            },
@@ -81,9 +83,11 @@ 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"             },
+    { LLM_ARCH_DREAM,            "dream"            },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -784,6 +788,36 @@ 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_PLAMO2,
+        {
+            { 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_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_SSM_IN,          "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,      "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_X,           "blk.%d.ssm_x" },
+            { LLM_TENSOR_SSM_DT,          "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,           "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,           "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
+            { LLM_TENSOR_SSM_DT_NORM,     "blk.%d.ssm_dt_norm" },
+            { LLM_TENSOR_SSM_B_NORM,      "blk.%d.ssm_b_norm" },
+            { LLM_TENSOR_SSM_C_NORM,      "blk.%d.ssm_c_norm" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+        },
+    },
     {
         LLM_ARCH_CODESHELL,
         {
@@ -1477,6 +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,
         {
@@ -1793,6 +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,
         {
@@ -1860,6 +1939,23 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
         },
     },
+    {
+        LLM_ARCH_DREAM,
+        {
+            { 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" },
+        },
+    },
 };
 
 static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
@@ -2094,6 +2190,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
     switch (arch) {
         case LLM_ARCH_JAMBA:
         case LLM_ARCH_FALCON_H1:
+        case LLM_ARCH_PLAMO2:
         case LLM_ARCH_GRANITE_HYBRID:
         case LLM_ARCH_LFM2:
             return true;
@@ -2101,3 +2198,12 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
             return false;
     }
 }
+
+bool llm_arch_is_diffusion(const llm_arch & arch) {
+    switch (arch) {
+        case LLM_ARCH_DREAM:
+            return true;
+        default:
+            return false;
+    }
+}

src/llama-arch.h

@@ -38,6 +38,7 @@ enum llm_arch {
     LLM_ARCH_PHI3,
     LLM_ARCH_PHIMOE,
     LLM_ARCH_PLAMO,
+    LLM_ARCH_PLAMO2,
     LLM_ARCH_CODESHELL,
     LLM_ARCH_ORION,
     LLM_ARCH_INTERNLM2,
@@ -71,6 +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,
@@ -85,9 +87,11 @@ 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,
+    LLM_ARCH_DREAM,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -478,3 +482,4 @@ const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
 
 bool llm_arch_is_recurrent(const llm_arch & arch);
 bool llm_arch_is_hybrid   (const llm_arch & arch);
+bool llm_arch_is_diffusion(const llm_arch & arch);

src/llama-batch.cpp

@@ -27,6 +27,7 @@ bool llama_batch_allocr::init(
         const llama_vocab & vocab,
         const llama_memory_i * memory,
         uint32_t n_embd,
+        uint32_t n_seq_max,
         bool output_all) {
     clear();
 
@@ -40,6 +41,11 @@ bool llama_batch_allocr::init(
     // validate input batch
     //
 
+    if (n_seq_max > LLAMA_MAX_SEQ) {
+        LLAMA_LOG_ERROR("%s: n_seq_max = %d > %d\n", __func__, n_seq_max, LLAMA_MAX_SEQ);
+        return false;
+    }
+
     if (batch.token) {
         for (int32_t i = 0; i < batch.n_tokens; ++i) {
             if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= vocab.n_tokens()) {
@@ -52,8 +58,8 @@ bool llama_batch_allocr::init(
     if (batch.seq_id) {
         for (int32_t i = 0; i < batch.n_tokens; ++i) {
             for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
-                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= LLAMA_MAX_SEQ)) {
-                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], LLAMA_MAX_SEQ);
+                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= (llama_seq_id) n_seq_max)) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], (llama_seq_id) n_seq_max);
                     return false;
                 }
             }
@@ -86,7 +92,7 @@ bool llama_batch_allocr::init(
 
         // initialize the starting position for each sequence based on the positions in the memory
         llama_pos p0[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             if (!memory) {
                 // if no memory -> start from 0
                 p0[s] = 0;
@@ -143,13 +149,16 @@ bool llama_batch_allocr::init(
     // compute stats
     //
 
-    this->n_embd = n_embd;
+    this->n_embd    = n_embd;
+    this->n_seq_max = n_seq_max;
 
     // count the outputs in this batch
     for (int32_t i = 0; i < batch.n_tokens; ++i) {
         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) {
@@ -189,7 +198,7 @@ bool llama_batch_allocr::init(
             seq_set_map[cur].push_back(i);
         }
 
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             if (seq_set_unq.test(s)) {
                 seq_idx[s] = seq_id_unq.size();
                 seq_id_unq.push_back(s);
@@ -201,7 +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,
@@ -214,6 +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);
@@ -241,7 +251,7 @@ bool llama_batch_allocr::init(
     // consistency checks
     //
 
-    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+    for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_pos[s].empty()) {
             continue;
         }
@@ -284,8 +294,8 @@ bool llama_batch_allocr::init(
     }
 
     if (memory) {
-        for (int32_t s0 = 0; s0 < LLAMA_MAX_SEQ; ++s0) {
-            for (int32_t s1 = 0; s1 < LLAMA_MAX_SEQ; ++s1) {
+        for (uint32_t s0 = 0; s0 < n_seq_max; ++s0) {
+            for (uint32_t s1 = 0; s1 < n_seq_max; ++s1) {
                 if (seq_cpl[s0][s1]) {
                     if (memory->seq_pos_min(s0) != memory->seq_pos_min(s1) ||
                         memory->seq_pos_max(s0) != memory->seq_pos_max(s1)) {
@@ -316,12 +326,12 @@ bool llama_batch_allocr::init(
     //
     {
         seq_set_t cur_seq_set[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             cur_seq_set[s].set();
         }
 
         llama_pos cur_seq_pos[LLAMA_MAX_SEQ];
-        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        for (uint32_t s = 0; s < n_seq_max; ++s) {
             cur_seq_pos[s] = -1;
         }
 
@@ -357,39 +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;
@@ -430,8 +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) {
@@ -646,78 +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 (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+    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);
     }
@@ -727,7 +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
@@ -48,6 +67,7 @@ public:
             const llama_vocab & vocab,
             const llama_memory_i * memory,
             uint32_t n_embd,
+            uint32_t n_seq_max,
             bool output_all);
 
     const llama_batch & get_batch() const;
@@ -100,6 +120,7 @@ private:
     const uint32_t n_pos_per_embd;
 
     uint32_t n_embd;
+    uint32_t n_seq_max;
     uint32_t n_outputs;
 
     std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
@@ -115,7 +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
@@ -135,20 +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          },
@@ -65,6 +66,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
     { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
     { "hunyuan-moe",       LLM_CHAT_TEMPLATE_HUNYUAN_MOE       },
+    { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -167,6 +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;
@@ -188,6 +193,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_DOTS1;
     } else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) {
         return LLM_CHAT_TEMPLATE_HUNYUAN_MOE;
+    } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
+        return LLM_CHAT_TEMPLATE_KIMI_K2;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -529,6 +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++) {
@@ -680,6 +703,26 @@ int32_t llm_chat_apply_template(
                 ss << "<|startoftext|>" << message->content << "<|extra_0|>";
             }
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_KIMI_K2) {
+        // moonshotai/Kimi-K2-Instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|im_system|>system<|im_middle|>";
+            } else if (role == "user") {
+                ss << "<|im_user|>user<|im_middle|>";
+            } else if (role == "assistant") {
+                ss << "<|im_assistant|>assistant<|im_middle|>";
+            } else if (role == "tool") {
+                ss << "<|im_system|>tool<|im_middle|>";
+            }
+
+            ss << message->content << "<|im_end|>";
+
+            if (add_ass) {
+                ss << "<|im_assistant|>assistant<|im_middle|>";
+            }
+        }
     } else {
         // template not supported
         return -1;

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,
@@ -45,6 +46,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_HUNYUAN_MOE,
+    LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

src/llama-context.cpp

@@ -98,10 +98,20 @@ llama_context::llama_context(
         LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
         cparams.n_batch = GGML_KQ_MASK_PAD;
     }
-
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
     cparams.op_offload = params.op_offload;
+    cparams.kv_unified = params.kv_unified;
+
+    {
+        const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
+        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__);
+            cparams.kv_unified = true;
+        }
+    }
 
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
@@ -112,6 +122,7 @@ llama_context::llama_context(
     LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
     LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
     LLAMA_LOG_INFO("%s: flash_attn    = %d\n",   __func__, cparams.flash_attn);
+    LLAMA_LOG_INFO("%s: kv_unified    = %s\n",   __func__, cparams.kv_unified ? "true" : "false");
     LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
     LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
 
@@ -227,8 +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
@@ -267,7 +278,7 @@ llama_context::llama_context(
 
     // reserve worst-case graph
     if (!hparams.vocab_only && memory) {
-        const uint32_t n_seqs = cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
@@ -300,7 +311,7 @@ llama_context::llama_context(
 
         // reserve with tg graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(1, 1, 1, mctx.get());
+            auto * gf = graph_reserve(n_seqs, n_seqs, n_seqs, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -311,6 +322,10 @@ llama_context::llama_context(
 
         // reserve again with pp graph to avoid ggml-alloc reallocations during inference
         {
+            // TODO: not sure if the following graph would be worster case for multi-stream KV caches:
+            //
+            // auto * gf = graph_reserve(n_tokens, 1, n_tokens, mctx.get());
+            //
             auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
@@ -388,10 +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;
 }
@@ -463,6 +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__);
         }
@@ -475,7 +491,7 @@ bool llama_context::kv_self_update(bool optimize) {
             throw std::runtime_error("failed to initialize memory context");
         }
 
-        const uint32_t n_seqs   = cparams.n_seq_max;
+        const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
         const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -678,38 +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;
@@ -731,16 +768,19 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd;
+    const int32_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
-    if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, true)) {
+    if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
 
     const uint32_t n_tokens = balloc->get_n_tokens();
 
+    // [TAG_NO_CACHE_PAD]
+    // TODO: add new split mode where we pad the input sequences so that ubatch.equal_seqs == true
     const llama_ubatch ubatch = balloc->split_simple(n_tokens);
 
     // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
@@ -767,9 +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
@@ -778,7 +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;
 
@@ -791,10 +828,20 @@ int llama_context::encode(const llama_batch & batch_inp) {
         }
     }
 
+    auto * t_logits = res->get_logits();
     auto * t_embd = res->get_embd_pooled() ? res->get_embd_pooled() : res->get_embd();
 
+    // extract logits
+   if (logits && t_logits) {
+        ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits);
+        GGML_ASSERT(backend_res != nullptr);
+        GGML_ASSERT(logits != nullptr);
+
+        ggml_backend_tensor_get_async(backend_res, t_logits, logits, 0, n_tokens*n_vocab*sizeof(float));
+    }
+
     // extract embeddings
-    if (t_embd) {
+    if (embd && t_embd) {
         ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd);
         GGML_ASSERT(backend_embd != nullptr);
 
@@ -844,10 +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;
@@ -899,7 +942,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     // when computing embeddings, all tokens are output
     const bool output_all = cparams.embeddings;
 
-    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, output_all)) {
+    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, output_all)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
         return -1;
     }
@@ -1005,11 +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
@@ -1190,10 +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;
 }
 
@@ -1275,20 +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) {
@@ -1301,6 +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;
@@ -1310,18 +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__);
@@ -1331,28 +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(
@@ -1930,6 +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;
 }
@@ -1938,6 +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;
 }
 
 //
@@ -2028,7 +2060,7 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, true)) {
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
             LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
             return;
         }
@@ -2064,8 +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;
             {
@@ -2187,6 +2224,7 @@ llama_context_params llama_context_default_params() {
         /*.no_perf                     =*/ true,
         /*.op_offload                  =*/ true,
         /*.swa_full                    =*/ true,
+        /*.kv_unified                  =*/ false,
     };
 
     return result;
@@ -2807,6 +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-cparams.h

@@ -11,8 +11,8 @@ struct llama_cparams {
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;
-    int      n_threads;       // number of threads to use for generation
-    int      n_threads_batch; // number of threads to use for batch processing
+    int32_t  n_threads;       // number of threads to use for generation
+    int32_t  n_threads_batch; // number of threads to use for batch processing
 
     float rope_freq_base;
     float rope_freq_scale;
@@ -33,6 +33,7 @@ struct llama_cparams {
     bool no_perf;
     bool warmup;
     bool op_offload;
+    bool kv_unified;
 
     enum llama_pooling_type pooling_type;
 

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,
@@ -982,13 +1143,16 @@ ggml_tensor * llm_graph_context::build_attn_mha(
              float     kq_scale) const {
     const bool v_trans = v->nb[1] > v->nb[2];
 
+    // split the batch into streams if needed
+    const auto n_stream = k->ne[3];
+
+    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);
+
     q = ggml_permute(ctx0, q, 0, 2, 1, 3);
     k = ggml_permute(ctx0, k, 0, 2, 1, 3);
     v = ggml_permute(ctx0, v, 0, 2, 1, 3);
 
-    const auto n_tokens = q->ne[1];
-    const auto n_head   = q->ne[2];
-    const auto n_kv     = k->ne[1];
+    const auto n_kv = k->ne[1];
 
     ggml_tensor * cur;
 
@@ -1030,7 +1194,7 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 #endif
         }
 
-        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
     } else {
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
 
@@ -1075,7 +1239,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
 
-        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
+        // recombine streams
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
 
         if (!cparams.offload_kqv) {
             // all nodes between the KV store and the attention output are run on the CPU
@@ -1102,7 +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,
@@ -1122,11 +1286,15 @@ 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());
+
     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) {
@@ -1156,13 +1324,14 @@ static std::unique_ptr<llm_graph_input_attn_kv_unified> build_attn_inp_kv_unifie
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
 
-        const auto n_kv = mctx_cur->get_n_kv();
+        const auto n_kv     = mctx_cur->get_n_kv();
         const auto n_tokens = ubatch.n_tokens;
+        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
 
         inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1181,7 +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,
@@ -1214,7 +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) {
@@ -1234,7 +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,
@@ -1281,7 +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) {
@@ -1314,7 +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,
@@ -1336,7 +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) {
@@ -1362,13 +1528,15 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
 
     auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
 
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
+
     {
         const auto n_kv = mctx_cur->get_base()->get_n_kv();
 
         inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1382,7 +1550,7 @@ llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unif
         inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1392,7 +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,
@@ -1450,21 +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;
@@ -1474,7 +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);
@@ -1514,7 +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 {
@@ -249,16 +265,18 @@ 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_kq_mask() const { return self_kq_mask_cnv; }
 
     ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -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; }
@@ -289,14 +309,14 @@ public:
     ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
     ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
     ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -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-hparams.cpp

@@ -65,6 +65,46 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
+bool llama_hparams::is_n_embd_k_gqa_variable() const {
+    const uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_k_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+bool llama_hparams::is_n_embd_v_gqa_variable() const {
+    const uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_v_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+uint32_t llama_hparams::n_embd_k_gqa_max() const {
+    uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_k_gqa(il));
+    }
+
+    return val;
+}
+
+uint32_t llama_hparams::n_embd_v_gqa_max() const {
+    uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_v_gqa(il));
+    }
+
+    return val;
+}
+
 uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models

src/llama-hparams.h

@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3
+#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
@@ -191,6 +191,14 @@ struct llama_hparams {
     // dimension of value embeddings across all k-v heads
     uint32_t n_embd_v_gqa(uint32_t il = 0) const;
 
+    // true if any layer has a different n_embd_k_gqa/n_embd_v_gqa
+    bool is_n_embd_k_gqa_variable() const;
+    bool is_n_embd_v_gqa_variable() const;
+
+    // return the maximum n_embd_k_gqa/n_embd_v_gqa across all layers
+    uint32_t n_embd_k_gqa_max() const;
+    uint32_t n_embd_v_gqa_max() const;
+
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
     uint32_t n_embd_r() const;