ggml: adds CONV_2D op and direct GEMM Vulkan implementation (#14316)

* ggml/ggml-vulkan/test-backend-ops: adds CONV_2D for Vulkan

* ggml-vulkan: adds f32 scalar shader to compute 2D convolution directly
with gemm (no need for im2col),

* test-backend-ops: adds test_case_ref to check the validity/performance of ops
against reference implementations having different graphs, adds tests

* * Performance fixes: minimized branch divergence, uses collectives to
  eliminate redundant calculation, macros removed.

* Kernel shared memory size check

* Updates test-backend-ops to support graphs for performance
  measurement.

* * Apple/Win32 compile errors fixed

* Subgroup size used to determine tile size -> fixes llvmpipe errors.

* Collectives disabled by default.

* Intel support is disabled as the performance is poor.

* Conv2d enabled for Intel with disabled collectives, disabled for Apple

* test-backend-ops modifications are reverted

* Trailing spaces and missing override fixed.

* Triggering pipeline relaunch.

* Code formatted with .clang-format.

.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/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -40,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         multiple: true
     validations:
       required: true

.github/labeler.yml

@@ -1,10 +1,4 @@
 # https://github.com/actions/labeler
-Kompute:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-kompute.h
-            - ggml/src/ggml-kompute/**
-            - README-kompute.md
 Apple Metal:
     - changed-files:
         - any-glob-to-any-file:

.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 $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
   ubuntu-cpu-cmake:
     strategy:
       matrix:
@@ -342,6 +405,72 @@ jobs:
           cd build
           export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
+          ctest -L main --verbose --timeout 4200
+
+  ubuntu-22-cmake-webgpu:
+    runs-on: ubuntu-22.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: ubuntu-22-cmake-webgpu
+          evict-old-files: 1d
+
+      - name: Vulkan SDK Dependencies
+        id: vulkan-depends
+        run: |
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
+          ARTIFACTS_JSON=$(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:
@@ -740,9 +869,6 @@ jobs:
           - build: 'llvm-arm64-opencl-adreno'
             arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
-         # - build: 'kompute-x64'
-         #   arch: 'x64'
-         #   defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
 
     steps:
       - name: Clone
@@ -756,12 +882,6 @@ jobs:
           variant: ccache
           evict-old-files: 1d
 
-      - name: Clone Kompute submodule
-        id: clone_kompute
-        if: ${{ matrix.build == 'kompute-x64' }}
-        run: |
-          git submodule update --init ggml/src/ggml-kompute/kompute
-
       - name: Download OpenBLAS
         id: get_openblas
         if: ${{ matrix.build == 'openblas-x64' }}
@@ -777,7 +897,7 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.build == 'vulkan-x64' }}
         run: |
           curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install

.github/workflows/release.yml

@@ -49,7 +49,8 @@ jobs:
         run: |
           sysctl -a
           cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
@@ -103,7 +104,8 @@ jobs:
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
           cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON
@@ -160,6 +162,8 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DGGML_BACKEND_DL=ON \
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
@@ -211,6 +215,8 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DGGML_BACKEND_DL=ON \
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \

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

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

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "kompute"]
-	path = ggml/src/ggml-kompute/kompute
-	url = https://github.com/nomic-ai/kompute.git

CMakeLists.txt

@@ -120,7 +120,6 @@ endfunction()
 
 llama_option_depr(FATAL_ERROR LLAMA_CUBLAS              GGML_CUDA)
 llama_option_depr(WARNING     LLAMA_CUDA                GGML_CUDA)
-llama_option_depr(WARNING     LLAMA_KOMPUTE             GGML_KOMPUTE)
 llama_option_depr(WARNING     LLAMA_METAL               GGML_METAL)
 llama_option_depr(WARNING     LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
 llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)

CMakePresets.json

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

README.md

@@ -6,9 +6,9 @@
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md)
 
-Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
+LLM inference in C/C++
 
 ## Recent API changes
 
@@ -17,10 +17,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
+- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
+- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
-- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
@@ -134,6 +133,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
+- [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
 
 #### Multimodal
 
@@ -269,6 +269,8 @@ 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

ci/run.sh

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

common/CMakeLists.txt

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

common/arg.cpp

@@ -1464,6 +1464,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.swa_full = true;
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
+    add_opt(common_arg(
+        {"--kv-unified", "-kvu"},
+        string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/14363)", params.kv_unified ? "true" : "false"),
+        [](common_params & params) {
+            params.kv_unified = true;
+        }
+    ).set_env("LLAMA_ARG_KV_SPLIT"));
     add_opt(common_arg(
         {"--no-context-shift"},
         string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"),
@@ -2734,6 +2742,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.public_path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
+    add_opt(common_arg(
+        {"--api-prefix"}, "PREFIX",
+        string_format("prefix path the server serves from, without the trailing slash (default: %s)", params.api_prefix.c_str()),
+        [](common_params & params, const std::string & value) {
+            params.api_prefix = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_API_PREFIX"));
     add_opt(common_arg(
         {"--no-webui"},
         string_format("Disable the Web UI (default: %s)", params.webui ? "enabled" : "disabled"),
@@ -3416,5 +3431,34 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
+    // diffusion parameters
+    add_opt(common_arg(
+        { "--diffusion-steps" }, "N",
+        string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
+        [](common_params & params, int value) { params.diffusion.steps = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-eps" }, "F",
+        string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
+        [](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-algorithm" }, "N",
+        string_format("diffusion algorithm: 0=ORIGIN, 1=MASKGIT_PLUS, 2=TOPK_MARGIN, 3=ENTROPY (default: %d)",
+                      params.diffusion.algorithm),
+        [](common_params & params, int value) { params.diffusion.algorithm = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-alg-temp" }, "F",
+        string_format("algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
+        [](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "--diffusion-visual" },
+        string_format("enable visual diffusion mode (show progressive generation) (default: %s)",
+                      params.diffusion.visual_mode ? "true" : "false"),
+        [](common_params & params) { params.diffusion.visual_mode = true; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+
     return ctx_arg;
 }

common/common.cpp

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

common/common.h

@@ -81,6 +81,7 @@ enum llama_example {
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
+    LLAMA_EXAMPLE_DIFFUSION,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -177,7 +178,8 @@ struct common_params_sampling {
     std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
 
-    std::vector<llama_logit_bias> logit_bias; // logit biases to apply
+    std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
+    std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
     // print the parameters into a string
     std::string print() const;
@@ -217,6 +219,14 @@ struct common_params_vocoder {
     bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };
 
+struct common_params_diffusion {
+    int32_t steps       = 64;     // number of diffusion steps
+    float   eps         = 1e-3f;  // epsilon for timesteps
+    int32_t algorithm   = 0;      // diffusion algorithm (0=ORIGIN, 1=MASKGIT_PLUS, 2=TOPK_MARGIN, 3=ENTROPY)
+    float   alg_temp    = 0.0f;   // algorithm temperature
+    bool    visual_mode = false;  // show progressive diffusion on screen
+};
+
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
@@ -268,6 +278,7 @@ struct common_params {
     struct common_params_sampling    sampling;
     struct common_params_speculative speculative;
     struct common_params_vocoder     vocoder;
+    struct common_params_diffusion   diffusion;
 
     struct common_params_model model;
 
@@ -330,6 +341,7 @@ struct common_params {
     bool no_perf           = false; // disable performance metrics
     bool ctx_shift         = true;  // context shift on inifinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+    bool kv_unified        = false; // enable unified KV cache
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // use mmap for faster loads
@@ -370,6 +382,7 @@ struct common_params {
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
+    std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
@@ -521,6 +534,7 @@ static bool string_starts_with(const std::string & str,
 
 // While we wait for C++20's std::string::ends_with...
 bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
+bool string_remove_suffix(std::string & str, const std::string_view & suffix);
 size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);

convert_hf_to_gguf_update.py

@@ -7,7 +7,6 @@ import pathlib
 import re
 
 import requests
-import sys
 import json
 import shutil
 import argparse
@@ -69,8 +68,7 @@ args = parser.parse_args()
 hf_token = args.hf_token if args.hf_token is not None else hf_token
 
 if hf_token is None:
-    logger.error("HF token is required. Please provide it as an argument or set it in ~/.cache/huggingface/token")
-    sys.exit(1)
+    logger.warning("HF token not found. You can provide it as an argument or set it in ~/.cache/huggingface/token")
 
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
@@ -128,6 +126,10 @@ models = [
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
+    {"name": "a.x-4.0",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/skt/A.X-4.0", },
+    {"name": "midm-2.0",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct", },
+    {"name": "lfm2",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2-Tokenizer"},
+    {"name": "exaone4",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-32B", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -137,11 +139,18 @@ pre_computed_hashes = [
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
     {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
     {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
+    {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
+    # falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
+    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
+    {"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)
@@ -222,7 +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")
@@ -230,11 +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
@@ -244,15 +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 +385,28 @@ 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 the official [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 done everything in the Vulkan SDK guide provided in the first step, run the following command to verify that everything is set up correctly:
 ```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
+```
+
+Finally, after finishing your build, you should be able to do this:
+```bash
 # 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
+./build/bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
 
 # 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 +553,23 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
+## WebGPU [In Progress]
+
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
+
+In the llama.cpp directory, build with CMake:
+
+```
+cmake -B build -DGGML_WEBGPU=ON
+cmake --build build --config Release
+```
+
+### Browser Support
+
+WebGPU allows cross-platform access to the GPU from supported browsers. We utilize [Emscripten](https://emscripten.org/) to compile ggml's WebGPU backend to WebAssembly. Emscripten does not officially support WebGPU bindings yet, but Dawn currently maintains its own WebGPU bindings called emdawnwebgpu.
+
+Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/src/emdawnwebgpu/) to download or build the emdawnwebgpu package (Note that it might be safer to build the emdawbwebgpu package locally, so that it stays in sync with the version of Dawn you have installed above). When building using CMake, the path to the emdawnwebgpu port file needs to be set with the flag `EMDAWNWEBGPU_DIR`.
+
 ## IBM Z & LinuxONE
 
 To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)

docs/development/HOWTO-add-model.md

@@ -83,20 +83,22 @@ NOTE: Tensor names must end with `.weight` or `.bias` suffixes, that is the conv
 
 ### 2. Define the model architecture in `llama.cpp`
 
-The model params and tensors layout must be defined in `llama.cpp`:
-1. Define a new `llm_arch`
-2. Define the tensors layout in `LLM_TENSOR_NAMES`
-3. Add any non-standard metadata in `llm_load_hparams`
-4. Create the tensors for inference in `llm_load_tensors`
-5. If the model has a RoPE operation, add the rope type in `llama_rope_type`
+The model params and tensors layout must be defined in `llama.cpp` source files:
+1. Define a new `llm_arch` enum value in `src/llama-arch.h`.
+2. In `src/llama-arch.cpp`:
+    - Add the architecture name to the `LLM_ARCH_NAMES` map.
+    - Add the tensor mappings to the `LLM_TENSOR_NAMES` map.
+3. Add any non-standard metadata loading in the `llama_model_loader` constructor in `src/llama-model-loader.cpp`.
+4. If the model has a RoPE operation, add a case for the architecture in `llama_model_rope_type` function in `src/llama-model.cpp`.
 
 NOTE: The dimensions in `ggml` are typically in the reverse order of the `pytorch` dimensions.
 
 ### 3. Build the GGML graph implementation
 
-This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `llama_build_graph`.
-
-Have a look at existing implementations like `build_llama`, `build_dbrx` or `build_bert`.
+This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `src/llama-model.cpp`.
+Create a new struct that inherits from `llm_graph_context` and implement the graph-building logic in its constructor.
+Have a look at existing implementations like `llm_build_llama`, `llm_build_dbrx` or `llm_build_bert`.
+Then, in the `llama_model::build_graph` method, add a case for your architecture to instantiate your new graph-building struct.
 
 Some `ggml` backends do not support all operations. Backend implementations can be added in a separate PR.
 

docs/ops.md

@@ -0,0 +1,95 @@
+# GGML Operations
+
+List of GGML operations and backend support status.
+
+Legend:
+- ✅ Fully supported by this backend
+- 🟡 Partially supported by this backend
+- ❌ Not supported by this backend
+
+| Operation | BLAS | CPU | CUDA | Metal |
+|-----------|------|------|------|------|
+|                              ABS | ❌ | ✅ | 🟡 | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ |
+|                              ADD | ❌ | ✅ | ✅ | 🟡 |
+|                             ADD1 | ❌ | ✅ | ✅ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ |
+|                            CLAMP | ❌ | ✅ | ✅ | 🟡 |
+|                           CONCAT | ❌ | ✅ | 🟡 | ✅ |
+|                             CONT | ❌ | ✅ | 🟡 | ✅ |
+|                       CONV_2D_DW | ❌ | ✅ | ✅ | ❌ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ |
+|                CONV_TRANSPOSE_2D | ❌ | ✅ | ✅ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | 🟡 |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ❌ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 |
+|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ✅ | ✅ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | 🟡 |
+|                              DIV | ❌ | ✅ | ✅ | 🟡 |
+|                              DUP | ❌ | ✅ | 🟡 | 🟡 |
+|                              ELU | ❌ | ✅ | ❌ | 🟡 |
+|                              EXP | ❌ | ✅ | 🟡 | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | ✅ | 🟡 | 🟡 |
+|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | 🟡 |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | 🟡 |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | 🟡 |
+|                             GELU | ❌ | ✅ | 🟡 | 🟡 |
+|                         GELU_ERF | ❌ | ✅ | 🟡 | 🟡 |
+|                       GELU_QUICK | ❌ | ✅ | 🟡 | 🟡 |
+|                         GET_ROWS | ❌ | ✅ | 🟡 | ✅ |
+|                    GET_ROWS_BACK | ❌ | 🟡 | 🟡 | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ |
+|                      HARDSIGMOID | ❌ | ✅ | 🟡 | ❌ |
+|                        HARDSWISH | ❌ | ✅ | 🟡 | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | 🟡 |
+|                          L2_NORM | ❌ | ✅ | ✅ | ✅ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ |
+|                              LOG | ❌ | ✅ | ✅ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ |
+|                              MUL | ❌ | ✅ | ✅ | 🟡 |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       MUL_MAT_ID | ❌ | ✅ | ✅ | ✅ |
+|                              NEG | ❌ | ✅ | 🟡 | 🟡 |
+|                             NORM | ❌ | ✅ | ✅ | 🟡 |
+|                   OPT_STEP_ADAMW | ❌ | ✅ | ✅ | ❌ |
+|                         OUT_PROD | 🟡 | 🟡 | 🟡 | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ❌ | ✅ |
+|                          POOL_2D | ❌ | ✅ | ✅ | ✅ |
+|                            REGLU | ❌ | ✅ | ✅ | 🟡 |
+|                             RELU | ❌ | ✅ | 🟡 | 🟡 |
+|                           REPEAT | ❌ | ✅ | 🟡 | ✅ |
+|                      REPEAT_BACK | ❌ | ✅ | ✅ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | 🟡 |
+|                    RMS_NORM_BACK | ❌ | ✅ | ✅ | ❌ |
+|                     RMS_NORM_MUL | ❌ | ✅ | ✅ | ✅ |
+|                             ROPE | ❌ | ✅ | ✅ | ✅ |
+|                        ROPE_BACK | ❌ | ✅ | ✅ | ❌ |
+|                        RWKV_WKV6 | ❌ | ✅ | ✅ | ✅ |
+|                        RWKV_WKV7 | ❌ | ✅ | ✅ | ✅ |
+|                            SCALE | ❌ | ✅ | ✅ | ✅ |
+|                              SET | ❌ | ✅ | ❌ | ✅ |
+|                         SET_ROWS | ❌ | 🟡 | ❌ | 🟡 |
+|                              SGN | ❌ | ✅ | 🟡 | ❌ |
+|                          SIGMOID | ❌ | ✅ | 🟡 | 🟡 |
+|                             SILU | ❌ | ✅ | 🟡 | 🟡 |
+|                        SILU_BACK | ❌ | ✅ | ✅ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | 🟡 |
+|                         SOFT_MAX | ❌ | ✅ | ✅ | ✅ |
+|                    SOFT_MAX_BACK | ❌ | 🟡 | 🟡 | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | 🟡 |
+|                             SQRT | ❌ | ✅ | ✅ | 🟡 |
+|                         SSM_CONV | ❌ | ✅ | ✅ | ✅ |
+|                         SSM_SCAN | ❌ | ✅ | ✅ | ✅ |
+|                             STEP | ❌ | ✅ | 🟡 | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | 🟡 |
+|                              SUM | ❌ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ |
+|                           SWIGLU | ❌ | ✅ | ✅ | 🟡 |
+|                             TANH | ❌ | ✅ | 🟡 | 🟡 |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ |
+|                          UPSCALE | ❌ | ✅ | ✅ | 🟡 |

examples/CMakeLists.txt

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

examples/diffusion/CMakeLists.txt

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

examples/diffusion/diffusion-cli.cpp

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

examples/embedding/embedding.cpp

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

examples/eval-callback/eval-callback.cpp

@@ -136,6 +136,11 @@ static bool run(llama_context * ctx, const common_params & params) {
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
+    if (tokens.empty()) {
+        LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
+        return false;
+    }
+
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;

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

@@ -181,7 +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_KOMPUTE                         "ggml: use Kompute"                               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)
@@ -266,12 +267,12 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cpp.h
     include/ggml-cuda.h
-    include/ggml-kompute.h
     include/ggml-opt.h
     include/ggml-metal.h
     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}")
@@ -360,6 +361,13 @@ write_basic_package_version_file(
     VERSION ${GGML_INSTALL_VERSION}
     COMPATIBILITY SameMajorVersion)
 
+target_compile_definitions(ggml-base PRIVATE
+    GGML_VERSION="${GGML_INSTALL_VERSION}"
+    GGML_COMMIT="${GGML_BUILD_COMMIT}"
+)
+message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
+message(STATUS "ggml commit:  ${GGML_BUILD_COMMIT}")
+
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)

ggml/include/ggml-kompute.h

@@ -1,50 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define GGML_KOMPUTE_MAX_DEVICES 16
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
-
-#ifdef __cplusplus
-}
-#endif

ggml/include/ggml-webgpu.h

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

ggml/include/ggml.h

@@ -495,7 +495,7 @@ extern "C" {
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
         GGML_OP_POOL_2D_BACK,
-        GGML_OP_UPSCALE, // nearest interpolate
+        GGML_OP_UPSCALE,
         GGML_OP_PAD,
         GGML_OP_PAD_REFLECT_1D,
         GGML_OP_ROLL,
@@ -557,6 +557,8 @@ extern "C" {
         GGML_GLU_OP_REGLU,
         GGML_GLU_OP_GEGLU,
         GGML_GLU_OP_SWIGLU,
+        GGML_GLU_OP_GEGLU_ERF,
+        GGML_GLU_OP_GEGLU_QUICK,
 
         GGML_GLU_OP_COUNT,
     };
@@ -646,6 +648,9 @@ extern "C" {
 
     // misc
 
+    GGML_API const char * ggml_version(void);
+    GGML_API const char * ggml_commit(void);
+
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
     GGML_API int64_t ggml_time_ms(void);
     GGML_API int64_t ggml_time_us(void);
@@ -1144,6 +1149,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_geglu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // A: n columns, r rows,
     // B: n columns, r rows,
     GGML_API struct ggml_tensor * ggml_glu_split(
@@ -1167,6 +1188,16 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_geglu_erf_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1266,6 +1297,19 @@ extern "C" {
             struct ggml_tensor  * a,
             float                 s);
 
+    // x = s * a + b
+    GGML_API struct ggml_tensor * ggml_scale_bias(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
+    GGML_API struct ggml_tensor * ggml_scale_bias_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 s,
+        float                 b);
+
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set(
             struct ggml_context * ctx,
@@ -1510,8 +1554,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a    [ne0, ne01, ne02, ne03]
+    // mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
+    //
+    // broadcast:
+    //   ne02 % ne12 == 0
+    //   ne03 % ne13 == 0
+    //
     // fused soft_max(a*scale + mask*(ALiBi slope))
-    // mask is optional
     // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
@@ -1974,11 +2024,17 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // q:    [n_embd_k, n_batch,     n_head,    1]
-    // k:    [n_embd_k, n_kv,        n_head_kv, 1]
-    // v:    [n_embd_v, n_kv,        n_head_kv, 1] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   1] !! permuted !!
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    //
+    // broadcast:
+    //   n_head % n_head_kv == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
+    //
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -2017,7 +2073,8 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * ids);
 
     // partition into non-overlapping windows with padding if needed
     // example:

ggml/src/CMakeLists.txt

@@ -365,12 +365,12 @@ ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
-ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
 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
@@ -61,10 +65,6 @@
 #include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
@@ -177,6 +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
@@ -189,9 +192,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
 #ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
 #endif
@@ -575,7 +575,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);
     ggml_backend_load_best("sycl", silent, dir_path);

ggml/src/ggml-backend.cpp

@@ -352,21 +352,6 @@ ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
 
 // backend copy
 
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
 void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
 

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -67,6 +67,7 @@
 #include <aclnnop/aclnn_pow.h>
 #include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
+#include <aclnnop/aclnn_zero.h>
 #include <float.h>
 
 #include <cmath>
@@ -804,10 +805,11 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
         nb[i] = nb[i - 1] * ne[i - 1];
     }
 
-    ggml_cann_async_memset(ctx, buffer, n_bytes, 0);
     aclTensor* zero =
         ggml_cann_create_tensor(buffer, type, type_size, ne, nb, dims);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, zero);
     return zero;
+    GGML_UNUSED(n_bytes);
 }
 
 /**

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

@@ -2086,6 +2086,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+#pragma message("TODO: implement F32, F16, BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                return false;
+            } break;
         case GGML_OP_CPY: {
             ggml_tensor *src = op->src[0];
             if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
@@ -2182,12 +2189,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_CLAMP:
         case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
@@ -2205,6 +2210,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_SCALE:
+            float bias;
+            memcpy(&bias, (float*)op->op_params + 1, sizeof(float));
+            return bias == 0.0f; // TODO: support bias != 0.0f
+        case GGML_OP_SOFT_MAX:
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
+            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
         case GGML_OP_FLASH_ATTN_EXT:{
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2227,6 +2240,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }

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

@@ -2172,6 +2172,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
                     {
                         n_tasks = n_threads;
                     } break;

ggml/src/ggml-cpu/ops.cpp

@@ -3614,6 +3614,292 @@ static void ggml_compute_forward_swiglu(
     }
 }
 
+// ggml_compute_forward_geglu_erf
+
+static void ggml_compute_forward_geglu_erf_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        float * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_erf_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            GGML_UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_erf_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_erf_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            GGML_UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_erf(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_geglu_erf_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_geglu_erf_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_geglu_quick
+
+static void ggml_compute_forward_geglu_quick_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        float * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_quick_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            GGML_UNUSED(x);
+            assert(!isnan(x));
+            assert(!isinf(x));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_quick_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int i1 = ir0; i1 < ir1; i1++) {
+        ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_quick_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#ifndef NDEBUG
+        for (int k = 0; k < nc; k++) {
+            const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
+            const float v = GGML_FP16_TO_FP32(x);
+            GGML_UNUSED(v);
+            assert(!isnan(v));
+            assert(!isinf(v));
+        }
+#endif
+    }
+}
+
+static void ggml_compute_forward_geglu_quick(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_geglu_quick_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_geglu_quick_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_norm
 
 static void ggml_compute_forward_norm_f32(
@@ -3729,6 +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);
             }
         }
@@ -4357,9 +4646,11 @@ static void ggml_compute_forward_scale_f32(
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    // scale factor
-    float v;
-    memcpy(&v, dst->op_params, sizeof(float));
+    float s; // scale factor
+    float b; // bias
+
+    memcpy(&s, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&b, (float *) dst->op_params + 1, sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -4378,12 +4669,22 @@ static void ggml_compute_forward_scale_f32(
 
     const size_t nb1 = dst->nb[1];
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        if (dst->data != src0->data) {
-            // src0 is same shape as dst => same indices
-            memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float));
+    if (b == 0.0f) {
+        for (int i1 = ir0; i1 < ir1; i1++) {
+            if (dst->data != src0->data) {
+                // src0 is same shape as dst => same indices
+                // TODO: add x parameter to ggml_vec_scale_f32 and remove this memcpy
+                memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float));
+            }
+            ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), s);
+        }
+    } else {
+        for (int i1 = ir0; i1 < ir1; i1++) {
+            ggml_vec_mad1_f32(nc,
+                (float *) ((char *) dst->data  + i1*nb1),
+                (float *) ((char *) src0->data + i1*nb1),
+                s, b);
         }
-        ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), v);
     }
 }
 
@@ -5232,14 +5533,17 @@ static void ggml_compute_forward_soft_max_f32(
     memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
     memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
-    // TODO: handle transposed/permuted matrices
-
     const int ith = params->ith;
     const int nth = params->nth;
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    //const int64_t ne11 = src1 ? src1->ne[1] : 1;
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
+
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
 
     // TODO: is this supposed to be ceil instead of floor?
     //       https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
@@ -5249,68 +5553,66 @@ static void ggml_compute_forward_soft_max_f32(
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    const int nc = src0->ne[0];
-    const int nr = ggml_nrows(src0);
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
+    float * wp = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        // ALiBi
-        const uint32_t h = (i1/ne01)%ne02; // head
-        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
+                const int64_t i11 = i01;
+                const int64_t i12 = i02%ne12;
+                const int64_t i13 = i03%ne13;
 
-        float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
-        float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+                // ALiBi
+                const uint32_t h = i02; // head
+                const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
 
-        // broadcast the mask across rows
-        ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
-        float       * mp_f32 = src1 ? (float       *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
+                float * sp = (float *)((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                float * dp = (float *)((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3);
 
-        ggml_vec_cpy_f32  (nc, wp, sp);
-        ggml_vec_scale_f32(nc, wp, scale);
-        if (mp_f32) {
-            if (use_f16) {
-                for (int i = 0; i < nc; ++i) {
-                    wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
+                // broadcast the mask across rows
+                ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
+                float       * mp_f32 = src1 ? (float       *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
+
+                ggml_vec_cpy_f32  (ne00, wp, sp);
+                ggml_vec_scale_f32(ne00, wp, scale);
+                if (mp_f32) {
+                    if (use_f16) {
+                        for (int i = 0; i < ne00; ++i) {
+                            wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
+                        }
+                    } else {
+                        for (int i = 0; i < ne00; ++i) {
+                            wp[i] += slope*mp_f32[i];
+                        }
+                    }
                 }
-            } else {
-                for (int i = 0; i < nc; ++i) {
-                    wp[i] += slope*mp_f32[i];
+
+#ifndef NDEBUG
+                for (int i = 0; i < ne00; ++i) {
+                    //printf("p[%d] = %f\n", i, p[i]);
+                    assert(!isnan(wp[i]));
                 }
+#endif
+
+                float max = -INFINITY;
+                ggml_vec_max_f32(ne00, &max, wp);
+
+                ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
+                assert(sum > 0.0);
+
+                sum = 1.0/sum;
+                ggml_vec_scale_f32(ne00, dp, sum);
+
+#ifndef NDEBUG
+                for (int i = 0; i < ne00; ++i) {
+                    assert(!isnan(dp[i]));
+                    assert(!isinf(dp[i]));
+                }
+#endif
             }
         }
-
-#ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
-            //printf("p[%d] = %f\n", i, p[i]);
-            assert(!isnan(wp[i]));
-        }
-#endif
-
-        float max = -INFINITY;
-        ggml_vec_max_f32(nc, &max, wp);
-
-        ggml_float sum = ggml_vec_soft_max_f32(nc, dp, wp, max);
-        assert(sum > 0.0);
-
-        sum = 1.0/sum;
-        ggml_vec_scale_f32(nc, dp, sum);
-
-#ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
-            assert(!isnan(dp[i]));
-            assert(!isinf(dp[i]));
-        }
-#endif
     }
 }
 
@@ -7766,7 +8068,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    ggml_type    const k_vec_dot_type      = ggml_get_type_traits_cpu(k->type)->vec_dot_type;
+    ggml_type         const k_vec_dot_type = ggml_get_type_traits_cpu(k->type)->vec_dot_type;
     ggml_from_float_t const q_to_vec_dot   = ggml_get_type_traits_cpu(k_vec_dot_type)->from_float;
     ggml_vec_dot_t    const kq_vec_dot     = ggml_get_type_traits_cpu(k->type)->vec_dot;
     ggml_to_float_t   const v_to_float     = ggml_get_type_traits(v->type)->to_float;
@@ -7798,7 +8100,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
             memset(VKQ32, 0, DV*sizeof(float));
         }
 
-        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL;
+        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]) : NULL;
 
         // k indices
         const int ik3 = iq3 / rk3;
@@ -8336,120 +8638,210 @@ void ggml_compute_forward_ssm_conv(
 static void ggml_compute_forward_ssm_scan_f32(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0]; // s
-    const ggml_tensor * src1 = dst->src[1]; // x
-    const ggml_tensor * src2 = dst->src[2]; // dt
-    const ggml_tensor * src3 = dst->src[3]; // A
-    const ggml_tensor * src4 = dst->src[4]; // B
-    const ggml_tensor * src5 = dst->src[5]; // C
+    const ggml_tensor * src0 = dst->src[0]; // s  {d_state, dim, n_head, n_seqs+}
+    const ggml_tensor * src1 = dst->src[1]; // x  {dim, n_head, n_seq_tokens, n_seqs}
+    const ggml_tensor * src2 = dst->src[2]; // dt {n_head, n_seq_tokens, n_seqs}
+    const ggml_tensor * src3 = dst->src[3]; // A  {d_state, n_head} or {1, n_head}
+    const ggml_tensor * src4 = dst->src[4]; // B  {d_state, n_group, n_seq_tokens, n_seqs}
+    const ggml_tensor * src5 = dst->src[5]; // C  {d_state, n_group, n_seq_tokens, n_seqs}
+    const ggml_tensor * src6 = dst->src[6]; // ids {n_seqs}
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int64_t nc  = src0->ne[0]; // d_state
-    const int64_t nr  = src0->ne[1]; // d_inner
-    const int64_t n_t = src1->ne[1]; // number of tokens per sequence
-    const int64_t n_s = src0->ne[2]; // number of sequences in the batch
+    const int64_t nc = src0->ne[0]; // d_state
+    const int64_t nr = src0->ne[1]; // dim
+    const int64_t nh = src1->ne[1]; // n_head
+    const int64_t ng = src4->ne[1];
+    const int64_t nt = src1->ne[2]; // number of tokens per sequence
+    const int64_t ns = src1->ne[3]; // number of sequences in the batch
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    // can't use ggml_nbytes because src1 is not necessarily contiguous
+    const int64_t s_off = ggml_nelements(src1) * ggml_element_size(src1);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*ns == ggml_nelements(dst));
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
     GGML_ASSERT(src2->nb[0] == sizeof(float));
     GGML_ASSERT(src3->nb[0] == sizeof(float));
     GGML_ASSERT(src4->nb[0] == sizeof(float));
     GGML_ASSERT(src5->nb[0] == sizeof(float));
-    // required for the dot product between s and C
-    GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float));
-    // required for per-sequence offsets for states
-    GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float));
-    // required to get correct offset for state destination (i.e. src1->nb[3])
-    GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float));
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
+    // allows optimizing the modulo since n_group should be a power of 2
+    GGML_ASSERT((ng & -ng) == ng);
 
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+    // heads per thread
+    const int dh = (nh + nth - 1)/nth;
 
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-    const int ir  = ir1 - ir0;
+    // head range for this thread
+    const int ih0 = dh*ith;
+    const int ih1 = MIN(ih0 + dh, nh);
 
-    #ifdef __ARM_FEATURE_SVE
-        for (int i3 = 0; i3 < n_s; ++i3) {
-            for (int i2 = 0; i2 < n_t; ++i2) {
-                const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
-                const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
-                const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
-                const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
-                const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
-                    float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                    float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
+    const int32_t * ids = (const int32_t *) src6->data;
 
-                // use the output as the source for the next token-wise iterations
-                if (i2 > 0) { s0 = s; }
+    for (int i3 = 0; i3 < ns; ++i3) {
+        const float * s0 = (const float *) ((const char *) src0->data + ids[i3]*(src0->nb[3])); // {d_state, dim, nh, ns}
+              float * s  = (      float *) ((      char *) dst->data  + i3*(src0->nb[3]) + s_off); // {d_state, dim, nh, ns}
 
-                // d_inner
-                for (int i1 = 0; i1 < ir; ++i1) {
-                    float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
-                    float x_dt = x[i1] * dt_soft_plus;
-                    svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
-                    svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
-                    svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
+        for (int i2 = 0; i2 < nt; ++i2) {
+            const float * x  = (const float *) ((const char *) src1->data + i2*(src1->nb[2]) + i3*(src1->nb[3])); // {dim, nh, nt, ns}
+            const float * dt = (const float *) ((const char *) src2->data + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {nh, nt, ns}
+            const float * A  = (const float *) ((const char *) src3->data); // {d_state, nh} or {1, nh}
+            const float * B  = (const float *) ((const char *) src4->data + i2*(src4->nb[2]) + i3*(src4->nb[3])); // {d_state, ng, nt, ns}
+            const float * C  = (const float *) ((const char *) src5->data + i2*(src5->nb[2]) + i3*(src5->nb[3])); // {d_state, ng, nt, ns}
+                  float * y  = (      float *) ((      char *) dst->data + i2*(nh*nr*sizeof(float)) + i3*(nt*nh*nr*sizeof(float))); // {dim, nh, nt, ns}
 
-                    for (int64_t k = 0; k < nc; k += svcntw()) {
-                        svfloat32_t vA = GGML_F32_VEC_LOAD(&A[i1*nc + k]);
-                        svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k]);
-                        svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k]);
-                        svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[i1*nc + k]);
+            if (src3->ne[0] == 1) {
+                // Mamba-2 has a scalar decay factor per head; dA can be outside the state-wise loop
 
-                        svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
-                        t1 = exp_ps_sve(svptrue_b32(), t1);
-                        svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+                    const float dA = expf(dt_soft_plus * A[h]);
 
-                        vs0 = GGML_F32_VEC_FMA(vs0, t1, t2);
-                        r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+                        float sumf = 0.0f;
+#if defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+                        const int ggml_f32_epr = svcntw();
+                        const int ggml_f32_step = 1 * ggml_f32_epr;
 
-                        GGML_F32_VEC_STORE(&s[i1*nc + k], vs0);
-                    }
-                    y[i1] = GGML_F32xt_REDUCE_ONE(r1_vector);
-                }
-            }
-        }
+                        const int np = (nc & ~(ggml_f32_step - 1));
+
+                        GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
+
+                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
+                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
+
+                        for (int i = 0; i < np; i += ggml_f32_step) {
+                            // TODO: maybe unroll more?
+                            for (int j = 0; j < 1; j++) {
+                                GGML_F32_VEC t0 = GGML_F32_VEC_LOAD(s0 + i + j*ggml_f32_epr + ii*nc);
+                                GGML_F32_VEC t1 = GGML_F32_VEC_LOAD(B + i + j*ggml_f32_epr + (h & (ng - 1))*nc);
+                                GGML_F32_VEC t2 = GGML_F32_VEC_LOAD(C + i + j*ggml_f32_epr + (h & (ng - 1))*nc);
+
+                                t0 = GGML_F32_VEC_MUL(t0, adA);
+                                t1 = GGML_F32_VEC_MUL(t1, axdt);
+
+                                t0 = GGML_F32_VEC_ADD(t0, t1);
+
+                                sum = GGML_F32_VEC_FMA(sum, t0, t2);
+
+                                GGML_F32_VEC_STORE(s + i + j*ggml_f32_epr + ii*nc, t0);
+                            }
+                        }
+
+                        sumf = GGML_F32xt_REDUCE_ONE(sum);
     #else
-        for (int i3 = 0; i3 < n_s; ++i3) {
-            for (int i2 = 0; i2 < n_t; ++i2) {
-                const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
-                const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
-                const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
-                const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
-                const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
-                    float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                    float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
+                        const int np = (nc & ~(GGML_F32_STEP - 1));
 
-                // use the output as the source for the next token-wise iterations
-                if (i2 > 0) { s0 = s; }
+                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
 
-                // d_inner
-                for (int i1 = 0; i1 < ir; ++i1) {
-                    // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78
-                    float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
-                    float x_dt = x[i1] * dt_soft_plus;
-                    float sumf = 0.0f;
-                    // d_state
-                    for (int i0 = 0; i0 < nc; ++i0) {
-                        int i = i0 + i1*nc;
-                        // state = prev_state * dA + dB * x
-                        float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt);
-                        // y = rowwise_dotprod(state, C)
-                        sumf += state * C[i0];
-                        s[i] = state;
+                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
+                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
+
+                        GGML_F32_VEC ax[GGML_F32_ARR];
+                        GGML_F32_VEC ay[GGML_F32_ARR];
+                        GGML_F32_VEC az[GGML_F32_ARR];
+
+                        for (int i = 0; i < np; i += GGML_F32_STEP) {
+                            for (int j = 0; j < GGML_F32_ARR; j++) {
+                                ax[j] = GGML_F32_VEC_LOAD(s0 + i + j*GGML_F32_EPR + ii*nc);
+                                ay[j] = GGML_F32_VEC_LOAD(B + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+                                az[j] = GGML_F32_VEC_LOAD(C + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+
+                                ax[j] = GGML_F32_VEC_MUL(ax[j], adA);
+                                ay[j] = GGML_F32_VEC_MUL(ay[j], axdt);
+
+                                ax[j] = GGML_F32_VEC_ADD(ax[j], ay[j]);
+
+                                sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], az[j]);
+
+                                GGML_F32_VEC_STORE(s + i + j*GGML_F32_EPR + ii*nc, ax[j]);
+                            }
+                        }
+
+                        // reduce sum0..sum3 to sum0
+                        GGML_F32_VEC_REDUCE(sumf, sum);
+    #endif
+#else
+                        const int np = 0;
+#endif
+                        // d_state
+                        for (int i0 = np; i0 < nc; ++i0) {
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
+                            // state = prev_state * dA + dB * x
+                            const float state = (s0[i] * dA) + (B[ig] * x_dt);
+                            // y = rowwise_dotprod(state, C)
+                            sumf += state * C[ig];
+                            s[i] = state;
+                        }
+                        y[ii] = sumf;
+                    }
+                }
+            } else {
+                // Mamba-1 has an element-wise decay factor for the states
+
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+#if defined(__ARM_FEATURE_SVE)
+                        svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
+                        svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
+                        svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
+
+                        // d_state
+                        // TODO: what happens when (d_state % svcntw()) != 0?
+                        for (int64_t k = 0; k < nc; k += svcntw()) {
+                            svfloat32_t vA = GGML_F32_VEC_LOAD(&A[h*nc + k]);
+                            svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k + (h & (ng - 1))*nc]);
+                            svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k + (h & (ng - 1))*nc]);
+                            svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[ii*nc + k]);
+
+                            svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
+                            t1 = exp_ps_sve(svptrue_b32(), t1);
+                            svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
+
+                            vs0 = GGML_F32_VEC_FMA(t2, vs0, t1);
+                            r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
+
+                            GGML_F32_VEC_STORE(&s[ii*nc + k], vs0);
+                        }
+                        y[ii] = GGML_F32xt_REDUCE_ONE(r1_vector);
+#else
+                        float sumf = 0.0f;
+                        // NOTE: can't really use GGML_SIMD here because d_state is usually 16
+                        //       and also because expf is used within the loop.
+                        // d_state
+                        for (int i0 = 0; i0 < nc; ++i0) {
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
+                            // state = prev_state * dA + dB * x
+                            const float state = (s0[i] * expf(dt_soft_plus * A[i0 + h*nc])) + (B[ig] * x_dt);
+                            // y = rowwise_dotprod(state, C)
+                            sumf += state * C[ig];
+                            s[i] = state;
+                        }
+                        y[ii] = sumf;
+#endif
                     }
-                    y[i1] = sumf;
                 }
             }
+            // use the output as the source when it's not the first token-wise iteration
+            s0 = s;
         }
-    #endif
+    }
 }
 
 void ggml_compute_forward_ssm_scan(
@@ -8688,6 +9080,14 @@ void ggml_compute_forward_glu(
             {
                 ggml_compute_forward_swiglu(params, dst);
             } break;
+        case GGML_GLU_OP_GEGLU_ERF:
+            {
+                ggml_compute_forward_geglu_erf(params, dst);
+            } break;
+        case GGML_GLU_OP_GEGLU_QUICK:
+            {
+                ggml_compute_forward_geglu_quick(params, dst);
+            } break;
         default:
             {
                 GGML_ABORT("fatal error");

ggml/src/ggml-cpu/simd-mappings.h

@@ -189,7 +189,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 #define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
 #define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, a, b, c)
+#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, b, c, a)
 #define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
 #define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)

ggml/src/ggml-cpu/vec.cpp

@@ -37,35 +37,35 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         for (int i = 0; i < np; i += ggml_f32_step) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
 
             ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
             ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2);
+            sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);
 
             ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
             ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3);
+            sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);
 
             ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
             ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4);
+            sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);
 
             ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
             ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5);
+            sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);
 
             ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
             ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6);
+            sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);
 
             ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
             ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7);
+            sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);
 
             ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
             ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8);
+            sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
         }
         // leftovers
         // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
@@ -73,7 +73,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         for (int i = np; i < np2; i += ggml_f32_epr) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
         }
         // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
         if (np2 < n) {
@@ -221,6 +221,9 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
     for (int i = np; i < n; ++i) {
         sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
     }
+
+    // if you hit this, you are likely running outside the FP range
+    assert(!isnan(sumf) && !isinf(sumf));
 #else
     for (int i = 0; i < n; ++i) {
         sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));

ggml/src/ggml-cpu/vec.h

@@ -163,49 +163,49 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
 
             GGML_F32_VEC_STORE(y + i, ay1);
 
             ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
             ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2);
+            ay2 = GGML_F32_VEC_FMA(ay2, ax2, vx);
 
             GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
 
             ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
             ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3);
+            ay3 = GGML_F32_VEC_FMA(ay3, ax3, vx);
 
             GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);
 
             ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
             ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4);
+            ay4 = GGML_F32_VEC_FMA(ay4, ax4, vx);
 
             GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);
 
             ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
             ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5);
+            ay5 = GGML_F32_VEC_FMA(ay5, ax5, vx);
 
             GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);
 
             ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
             ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6);
+            ay6 = GGML_F32_VEC_FMA(ay6, ax6, vx);
 
             GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);
 
             ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
             ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7);
+            ay7 = GGML_F32_VEC_FMA(ay7, ax7, vx);
 
             GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);
 
             ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
             ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8);
+            ay8 = GGML_F32_VEC_FMA(ay8, ax8, vx);
 
             GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
         }
@@ -215,7 +215,7 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
         for (int i = np; i < np2; i += ggml_f32_epr) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
 
             GGML_F32_VEC_STORE(y + i, ay1);
         }
@@ -351,6 +351,45 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
 #endif
 }
 
+inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, const float s, const float b) {
+#if defined(GGML_USE_ACCELERATE)
+    vDSP_vsmsa(x, 1, &s, &b, y, 1, n);
+#elif defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+        // scalar ; TODO: Write SVE code
+        for (int i = 0; i < n; ++i) {
+            y[i] = x[i]*s + b;
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
+
+        GGML_F32_VEC vs = GGML_F32_VEC_SET1(s);
+        GGML_F32_VEC vb = GGML_F32_VEC_SET1(b);
+
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ay[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_FMA(ay[j], vs, vb);
+
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] = x[i]*s + b;
+        }
+    #endif
+#else
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = x[i]*s + b;
+    }
+#endif
+}
+
 //inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) { for (int i = 0; i < n; ++i) y[i] *= v;          }
 inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #if defined(GGML_USE_ACCELERATE)
@@ -959,6 +998,46 @@ inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_
     }
 }
 
+inline static void ggml_vec_geglu_erf_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = x[i];
+        y[i] = 0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * g[i];
+    }
+}
+
+inline static void ggml_vec_geglu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+        float gi = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * gi);
+    }
+}
+
+#ifdef GGML_GELU_QUICK_FP16
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    uint16_t t;
+    for (int i = 0; i < n; ++i) {
+        ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+        memcpy(&t, &fp16, sizeof(uint16_t));
+        y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]) * g[i];
+    }
+}
+#else
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
+    }
+}
+#endif
+
+inline static void ggml_vec_geglu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[i16[i]]) * v);
+    }
+}
+
 inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
     ggml_float sum = 0.0;

ggml/src/ggml-cuda/common.cuh

@@ -175,6 +175,23 @@ static const char * cu_get_error_str(CUresult err) {
 #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
 #endif
 
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes)                                                       \
+        do {                                                                                                   \
+            static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = { false };                         \
+            const int   id                                                = ggml_cuda_get_device();            \
+            if (!shared_memory_limit_raised[id]) {                                                             \
+                CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes)); \
+                shared_memory_limit_raised[id] = true;                                                         \
+            }                                                                                                  \
+        } while (0)
+#else
+#    define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) \
+        do {                                             \
+            GGML_UNUSED(nbytes);                         \
+        } while (0)
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else

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

@@ -0,0 +1,251 @@
+#pragma once
+
+#include "ggml-common.h"
+
+static __device__ __forceinline__ void convert_f32_f32(const float * src, float * dst) {
+    *dst = *src;
+}
+
+static __device__ __forceinline__ void convert_f32_f16(const float * src, half * dst) {
+    *dst = __float2half(*src);
+}
+
+static __device__ __forceinline__ void convert_f32_bf16(const float * src, nv_bfloat16 * dst) {
+    *dst = *src;
+}
+
+static __device__ __forceinline__ void convert_f16_f16(const half * src, half * dst) {
+    *dst = *src;
+}
+
+static __device__ __forceinline__ void convert_f16_f32(const half * src, float * dst) {
+    *dst = *src;
+}
+
+static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static __device__ void quantize_f32_q4_0_block(const float * __restrict__ x, block_q4_0 * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = x[0       + j]*id;
+        const float x1 = x[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
+
+        y->qs[j]  = xi0;
+        y->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void quantize_f32_q4_1_block(const float * __restrict__ x, block_q4_1 * __restrict__ y) {
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = x[j];
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->dm.x = d;
+    y->dm.y = vmin;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (x[0       + j] - vmin)*id;
+        const float x1 = (x[QK4_1/2 + j] - vmin)*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
+
+        y->qs[j]  = xi0;
+        y->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void quantize_f32_q5_0_block(const float * __restrict__ x, block_q5_0 * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK5_0; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = x[0       + j]*id;
+        const float x1 = x[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
+
+        y->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+    memcpy(y->qh, &qh, sizeof(qh));
+}
+
+static __device__ void quantize_f32_q5_1_block(const float * __restrict__ x, block_q5_1 * __restrict__ y) {
+    float min = x[0];
+    float max = x[0];
+
+    for (int j = 1; j < QK5_1; ++j) {
+        const float v = x[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d  = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->dm.x = d;
+    y->dm.y = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (x[0       + j] - min)*id;
+        const float x1 = (x[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        y->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+    memcpy(y->qh, &qh, sizeof(qh));
+}
+
+static __device__ void quantize_f32_q8_0_block(const float * __restrict__ x, block_q8_0 * __restrict__ y) {
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = x[j];
+        amax = fmaxf(amax, fabsf(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    y->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = x[j]*id;
+        y->qs[j] = roundf(x0);
+    }
+}
+
+static __device__ void quantize_f32_iq4_nl_block(const float * __restrict__ x, block_iq4_nl * __restrict__ y) {
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_NL; ++j) {
+        const float v = x[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    float d = vmax / kvalues_iq4nl[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = x[0        + j]*id;
+        const float x1 = x[QK4_NL/2 + j]*id;
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
+        y->qs[j] = xi0 | (xi1 << 4);
+        const float v0 = kvalues_iq4nl[xi0];
+        const float v1 = kvalues_iq4nl[xi1];
+        const float w0 = x[0        + j]*x[0        + j];
+        const float w1 = x[QK4_NL/2 + j]*x[QK4_NL/2 + j];
+        sumqx += w0*v0*x[j] + w1*v1*x[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+    }
+
+    y->d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
+// Wrapper functions for cpy.cu compatibility
+static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    quantize_f32_q4_0_block((const float *)cxi, (block_q4_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    quantize_f32_q4_1_block((const float *)cxi, (block_q4_1 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
+    quantize_f32_q5_0_block((const float *)cxi, (block_q5_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
+    quantize_f32_q5_1_block((const float *)cxi, (block_q5_1 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    quantize_f32_q8_0_block((const float *)cxi, (block_q8_0 *)cdsti);
+}
+
+static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
+    quantize_f32_iq4_nl_block((const float *)cxi, (block_iq4_nl *)cdsti);
+}
+
+static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
+    convert_f32_f32((const float *)cxi, (float *)cdsti);
+}
+
+static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
+    convert_f32_f16((const float *)cxi, (half *)cdsti);
+}
+
+static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
+    convert_f32_bf16((const float *)cxi, (nv_bfloat16 *)cdsti);
+}
+
+static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
+    convert_f16_f16((const half *)cxi, (half *)cdsti);
+}
+
+static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
+    convert_f16_f32((const half *)cxi, (float *)cdsti);
+}

ggml/src/ggml-cuda/cpy.cu

@@ -1,46 +1,12 @@
 #include "cpy.cuh"
 #include "dequantize.cuh"
+#include "cpy-utils.cuh"
 #ifdef GGML_USE_MUSA
 #include "ggml-musa/mudnn.cuh"
 #endif // GGML_USE_MUSA
 
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
-static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    float * dsti = (float *) cdsti;
-
-    *dsti = *xi;
-}
-
-static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti;
-
-    *dsti = *xi;
-}
-
-static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    half * dsti = (half *) cdsti;
-
-    *dsti = __float2half(*xi);
-}
-
-static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
-    const half * xi = (const half *) cxi;
-    half * dsti = (half *) cdsti;
-
-    *dsti = *xi;
-}
-
-static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
-    const half * xi = (const half *) cxi;
-    float * dsti = (float *) cdsti;
-
-    *dsti = *xi;
-}
-
 template <cpy_kernel_t cpy_1>
 static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne,
                                    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@@ -71,29 +37,6 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const in
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
-static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q8_0 * dsti = (block_q8_0 *) cdsti;
-
-    float amax = 0.0f; // absolute max
-
-    for (int j = 0; j < QK8_0; j++) {
-        const float v = xi[j];
-        amax = fmaxf(amax, fabsf(v));
-    }
-
-    const float d = amax / ((1 << 7) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK8_0; ++j) {
-        const float x0 = xi[j]*id;
-
-        dsti->qs[j] = roundf(x0);
-    }
-}
-
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
 
@@ -106,139 +49,6 @@ static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     }
 }
 
-static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_0 * dsti = (block_q4_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_0; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -8;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK4_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK4_0/2 + j]*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_1 * dsti = (block_q4_1 *) cdsti;
-
-    float vmin = FLT_MAX;
-    float vmax = -FLT_MAX;
-
-    for (int j = 0; j < QK4_1; ++j) {
-        const float v = xi[j];
-
-        if (v < vmin) vmin = v;
-        if (v > vmax) vmax = v;
-    }
-
-    const float d  = (vmax - vmin) / ((1 << 4) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->dm.x = d;
-    dsti->dm.y = vmin;
-
-    for (int j = 0; j < QK4_1/2; ++j) {
-        const float x0 = (xi[0       + j] - vmin)*id;
-        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q5_0 * dsti = (block_q5_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK5_0; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -16;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK5_0/2 + j]*id;
-
-        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
-        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
-
-        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
-static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q5_1 * dsti = (block_q5_1 *) cdsti;
-
-    float min = xi[0];
-    float max = xi[0];
-
-    for (int j = 1; j < QK5_1; ++j) {
-        const float v = xi[j];
-        min = v < min ? v : min;
-        max = v > max ? v : max;
-    }
-
-    const float d  = (max - min) / 31;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->dm.x = d;
-    dsti->dm.y = min;
-
-    uint32_t qh = 0;
-    for (int j = 0; j < QK5_1/2; ++j) {
-        const float x0 = (xi[0       + j] - min)*id;
-        const float x1 = (xi[QK5_1/2 + j] - min)*id;
-
-        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
-    }
-    memcpy(dsti->qh, &qh, sizeof(qh));
-}
-
 template<dequantize_kernel_t dequant, int qk>
 static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
@@ -252,53 +62,6 @@ static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) {
     }
 }
 
-static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
-    }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
-}
-
-static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_NL; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    float d = vmax / kvalues_iq4nl[0];
-    const float id = d ? 1.0f/d : 0.0f;
-
-    float sumqx = 0, sumq2 = 0;
-    for (int j = 0; j < QK4_NL/2; ++j) {
-        const float x0 = xi[0        + j]*id;
-        const float x1 = xi[QK4_NL/2 + j]*id;
-        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
-        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
-        dsti->qs[j] = xi0 | (xi1 << 4);
-        const float v0 = kvalues_iq4nl[xi0];
-        const float v1 = kvalues_iq4nl[xi1];
-        const float w0 = xi[0        + j]*xi[0        + j];
-        const float w1 = xi[QK4_NL/2 + j]*xi[QK4_NL/2 + j];
-        sumqx += w0*v0*xi[j] + w1*v1*xi[QK4_NL/2 + j];
-        sumq2 += w0*v0*v0 + w1*v1*v1;
-    }
-
-    dsti->d = sumq2 > 0 ? sumqx/sumq2 : d;
-}
-
 template <cpy_kernel_t cpy_blck, int qk>
 static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne,
                                  const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,

ggml/src/ggml-cuda/cross-entropy-loss.cu

@@ -123,13 +123,7 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
     ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
 
     if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_f32<true>), smpbo);
         cross_entropy_loss_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
     } else {
         cross_entropy_loss_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
@@ -175,13 +169,7 @@ void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_ten
     const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
     if (nbytes_shared <= smpbo) {
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_back_f32<true>), smpbo);
         cross_entropy_loss_back_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
     } else {
         cross_entropy_loss_back_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);

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

@@ -32,7 +32,11 @@ typedef void (* fattn_kernel_t)(
         const int ne12,
         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,
@@ -519,7 +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;
@@ -533,8 +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;
@@ -543,14 +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;
@@ -569,7 +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;
@@ -615,16 +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();
@@ -642,11 +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]]
@@ -703,8 +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();
@@ -851,7 +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->nb[1] : 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,
@@ -866,11 +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

@@ -1223,7 +1223,11 @@ static __global__ void flash_attn_ext_f16(
         const int ne12,
         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,
@@ -1272,8 +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).
@@ -1283,17 +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 half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
-        float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+        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 + 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;
@@ -1322,17 +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 half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
-    float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
+    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 + 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;
@@ -1348,8 +1356,8 @@ static __global__ void flash_attn_ext_f16(
     GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
     GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
-    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
     GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
     GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
     GGML_UNUSED(ne2); GGML_UNUSED(ne3);

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

@@ -6,7 +6,7 @@
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
@@ -30,7 +30,11 @@ static __global__ void flash_attn_tile_ext_f16(
         const int ne12,
         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,
@@ -60,15 +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 + ne11*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.");
@@ -253,6 +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;
@@ -264,21 +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
@@ -288,8 +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(nb31); 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

@@ -6,7 +6,7 @@
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
@@ -30,7 +30,11 @@ static __global__ void flash_attn_tile_ext_f32(
         const int ne12,
         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,
@@ -58,8 +62,8 @@ static __global__ void flash_attn_tile_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(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
+        GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
         GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
         GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
         GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
@@ -72,15 +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 + ne11*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.");
 
@@ -263,6 +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;
@@ -274,22 +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
@@ -297,14 +305,14 @@ static __global__ void flash_attn_tile_ext_f32(
     GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
     GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
     GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap);
-    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02);
-    GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11);
-    GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12);
-    GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22);
-    GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03);
+    GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13);
+    GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32);
+    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

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

@@ -27,7 +27,11 @@ static __global__ void flash_attn_vec_ext_f16(
         const int ne12,
         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,
@@ -63,14 +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 + ne11*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.");
@@ -328,12 +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);
@@ -342,8 +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(nb31); 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

@@ -27,7 +27,11 @@ static __global__ void flash_attn_vec_ext_f32(
         const int ne12,
         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,
@@ -51,8 +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(nb31); 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);
@@ -75,13 +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
-    const half * maskh = (const half   *)  mask + ne11*ic0;
+    Q += nb03*sequence + nb02* head              + nb01*ic0;
+    K += nb13*sequence + nb12*(head / gqa_ratio);
+    V += nb23*sequence + nb22*(head / gqa_ratio);
 
-    const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
+    const half * maskh = (const half *) (mask + nb33*(sequence % ne33) + nb31*ic0);
+
+    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;
@@ -323,24 +330,25 @@ 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);
     GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale);
     GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1);
-    GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00);
-    GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10);
-    GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31);
-    GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
-    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21);
-    GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1);
-    GGML_UNUSED(ne2); GGML_UNUSED(ne3);
+    GGML_UNUSED(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(ne33);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); GGML_UNUSED(nb33);
+    GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03);
+    GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13);
+    GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23);
+    GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3);
     NO_DEVICE_CODE;
 #endif // FLASH_ATTN_AVAILABLE
 }

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

@@ -46,7 +46,11 @@ static __global__ void flash_attn_ext_f16(
         const int ne12,
         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,
@@ -93,17 +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 + (nb31/sizeof(half))* ic0;
-    const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
+    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);
 
@@ -398,7 +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) {
@@ -407,6 +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;
@@ -417,7 +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) {
@@ -431,7 +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);
@@ -440,7 +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(nb31); 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/getrows.cu

@@ -168,6 +168,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
             get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_I32:
+            get_rows_cuda_float((const int32_t *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
         case GGML_TYPE_BF16:
             get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
@@ -210,6 +214,10 @@ void get_rows_cuda(
             ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_I32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (int32_t *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
         case GGML_TYPE_F16:
             ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);

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

@@ -43,6 +43,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set-rows.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -2230,6 +2231,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_GET_ROWS_BACK:
             ggml_cuda_op_get_rows_back(ctx, dst);
             break;
+        case GGML_OP_SET_ROWS:
+            ggml_cuda_op_set_rows(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2299,6 +2303,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_EXP:
                     ggml_cuda_op_exp(ctx, dst);
                     break;
+                case GGML_UNARY_OP_ELU:
+                    ggml_cuda_op_elu(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2314,6 +2321,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_GLU_OP_SWIGLU:
                     ggml_cuda_op_swiglu(ctx, dst);
                     break;
+                case GGML_GLU_OP_GEGLU_ERF:
+                    ggml_cuda_op_geglu_erf(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    ggml_cuda_op_geglu_quick(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2577,6 +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];
 
@@ -2598,9 +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]);
@@ -3106,6 +3125,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_ELU:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -3116,6 +3136,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_GLU_OP_REGLU:
                 case GGML_GLU_OP_GEGLU:
                 case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
                     return ggml_is_contiguous_1(op->src[0]);
                 default:
                     return false;
@@ -3192,6 +3214,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F16:
                     case GGML_TYPE_F32:
+                    case GGML_TYPE_BF16:
+                    case GGML_TYPE_I32:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -3206,6 +3230,14 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                return (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
+                       op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q5_0 ||
+                       op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_IQ4_NL) &&
+                       op->src[0]->type == GGML_TYPE_F32 &&
+                       op->src[1]->type == GGML_TYPE_I64;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
@@ -3321,12 +3353,26 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
         case GGML_OP_LOG:
-        case GGML_OP_SSM_SCAN:
-        case GGML_OP_SSM_CONV:
             return true;
+        case GGML_OP_SSM_SCAN: {
+            if (op->src[3]->ne[0] == 1) {
+                // Mamba2
+                // (kernel only supports (d_state == 128 || d_state == 256) && d_head % 16 == 0)
+                return (op->src[0]->ne[0] == 128 || op->src[0]->ne[0] == 256) && op->src[0]->ne[1] % 16 == 0;
+            } else {
+                // Mamba
+                // (kernel only supports d_state == 16, d_head == 1, n_head % 128 == 0, n_group == 1)
+                return op->src[0]->ne[0] == 16 && op->src[0]->ne[1] == 1 && op->src[0]->ne[2] % 128 == 0 && op->src[4]->ne[1] == 1;
+            }
+        }
+        case GGML_OP_SSM_CONV: {
+            // assumes d_inner % threads == 0
+            return op->src[0]->ne[1] % 128 == 0;
+        }
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
         case GGML_OP_DIAG_MASK_INF:
+            return true;
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_SOFT_MAX_BACK: {
@@ -3351,7 +3397,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
         case GGML_OP_PAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -3375,9 +3420,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
-            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;
             }
@@ -3390,6 +3432,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
                 return true;
             }
+            if (op->src[3] && op->src[3]->ne[2] != 1) {
+                return false;
+            }
             return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) &&
                 op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
         }

ggml/src/ggml-cuda/mmq.cuh

@@ -3016,14 +3016,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-    static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-    if (!shared_memory_limit_raised[id]) {
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        shared_memory_limit_raised[id] = true;
-    }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, false>), nbytes_shared);
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, true>),  nbytes_shared);
 
     const int nty  = (args.nrows_x   + mmq_y - 1) / mmq_y;
     const int ntx  = (args.ncols_dst + mmq_x - 1) / mmq_x;

ggml/src/ggml-cuda/rope.cu

@@ -50,21 +50,19 @@ static __global__ void rope_norm(
 
     const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (i0 >= n_dims) {
-        const int i = row_dst*ne0 + i0;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
-        return;
-    }
-
     const int row_x     = row_dst % ne1;
     const int channel_x = row_dst / ne1;
 
     const int idst = row_dst*ne0 + i0;
     const int ix   = channel_x*s2 + row_x*s1 + i0;
 
+    if (i0 >= n_dims) {
+        dst[idst + 0] = x[ix + 0];
+        dst[idst + 1] = x[ix + 1];
+
+        return;
+    }
+
     const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
@@ -94,21 +92,19 @@ static __global__ void rope_neox(
 
     const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (i0 >= n_dims) {
-        const int i = row_dst*ne0 + i0;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
-        return;
-    }
-
     const int row_x     = row_dst % ne1;
     const int channel_x = row_dst / ne1;
 
     const int idst = row_dst*ne0 + i0/2;
     const int ix   = channel_x*s2 + row_x*s1 + i0/2;
 
+    if (i0 >= n_dims) {
+        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
+        dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];
+
+        return;
+    }
+
     const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
@@ -138,21 +134,19 @@ static __global__ void rope_multi(
 
     const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (i0 >= n_dims) {
-        const int i = row_dst*ne0 + i0;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
-        return;
-    }
-
     const int row_x     = row_dst % ne1;
     const int channel_x = row_dst / ne1;
 
     const int idst = row_dst*ne0 + i0/2;
     const int ix   = channel_x*s2 + row_x*s1 + i0/2;
 
+    if (i0 >= n_dims) {
+        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
+        dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];
+
+        return;
+    }
+
     const int sect_dims = sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
     const int sec_w = sections.v[1] + sections.v[0];
     const int sector = (i0 / 2) % sect_dims;

ggml/src/ggml-cuda/scale.cu

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

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

@@ -0,0 +1,288 @@
+#include "set-rows.cuh"
+#include "cpy-utils.cuh"
+
+typedef void (*set_rows_kernel_t)(const char * src, char * dst);
+
+template<typename src_t, typename dst_t>
+__device__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
+    GGML_UNUSED(src_f);
+    GGML_UNUSED(dst_f);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, half>(const float * src_f, half * dst_h) {
+    convert_f32_f16(src_f, dst_h);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, nv_bfloat16>(const float * src_f, nv_bfloat16 * dst_b) {
+    convert_f32_bf16(src_f, dst_b);
+}
+
+template<>
+__device__ __forceinline__ void set_rows_1<float, float>(const float * src_f, float * dst_f) {
+    convert_f32_f32(src_f, dst_f);
+}
+
+// Generic quantized set_rows kernel template
+template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+static __global__ void k_set_rows_quant(
+        const float * __restrict__ src0, const int64_t * __restrict__ src1, block_type * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s10, const int64_t s11, const int64_t s12,
+        const int64_t s1, const int64_t s2, const int64_t s3) {
+
+    const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
+    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
+
+    if (i >= ne_total) {
+        return;
+    }
+
+    const int64_t i_base = i * qk;
+    const int64_t i03 = i_base / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i_base - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+
+    const int64_t i12 = i03 % ne12;
+    const int64_t i11 = i02 % ne11;
+    const int64_t i10 = i01;
+
+    const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
+
+    const float * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
+    block_type * dst_row_ptr = dst + (dst_row*s1 + i02*s2 + i03*s3) / sizeof(block_type);
+
+    const float * src_block = src0_row + i00;
+    block_type * dst_block = dst_row_ptr + i00 / qk;
+
+    quantize_func(src_block, dst_block);
+}
+
+// Template dispatch function for quantized set_rows
+template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+static void set_rows_cuda_quant(
+        const float * src0_d, const int64_t * src1_d, block_type * dst_d,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+
+    GGML_ASSERT(ne00 % qk == 0);
+    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
+    const int num_blocks = (ne_total + CUDA_SET_ROWS_BLOCK_SIZE - 1) / CUDA_SET_ROWS_BLOCK_SIZE;
+    const dim3 block_size(CUDA_SET_ROWS_BLOCK_SIZE);
+    const dim3 grid_size(num_blocks);
+
+    const int64_t s01 = nb01/sizeof(float);
+    const int64_t s02 = nb02/sizeof(float);
+    const int64_t s03 = nb03/sizeof(float);
+    const int64_t s10 = nb10/sizeof(int64_t);
+    const int64_t s11 = nb11/sizeof(int64_t);
+    const int64_t s12 = nb12/sizeof(int64_t);
+    const int64_t s1  = nb1;
+    const int64_t s2  = nb2;
+    const int64_t s3  = nb3;
+
+    if (ne_total > 0) {
+        k_set_rows_quant<block_type, qk, quantize_func><<<grid_size, block_size, 0, stream>>>(
+            src0_d, src1_d, dst_d,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            s01, s02, s03,
+            s10, s11, s12,
+            s1, s2, s3);
+    }
+}
+
+template<typename src_t, typename dst_t>
+static __global__ void k_set_rows(
+        const src_t * __restrict__ src0, const int64_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t s10, const int64_t s11, const int64_t s12,
+        const int64_t s1, const int64_t s2, const int64_t s3) {
+
+    const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
+    const int64_t ne_total = ne00 * ne01 * ne02 * ne03;
+
+    if (i >= ne_total) {
+        return;
+    }
+
+    const int64_t i03 = i / (ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
+    const int64_t i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+
+    const int64_t i12 = i03 % ne12;
+    const int64_t i11 = i02 % ne11;
+    const int64_t i10 = i01;
+
+    const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
+
+    const src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
+    dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
+
+    const src_t* src_elem = src0_row + i00;
+    dst_t* dst_elem = dst_row_ptr + i00;
+    set_rows_1(src_elem, dst_elem);
+
+    GGML_UNUSED(ne10);
+    GGML_UNUSED(ne13);
+}
+
+template<typename src_t, typename dst_t>
+static void set_rows_cuda(
+        const src_t * src0_d, const int64_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+
+    const int64_t ne_total = ne00 * ne01 * ne02 * ne03;
+    const int num_blocks = (ne_total + CUDA_SET_ROWS_BLOCK_SIZE - 1) / CUDA_SET_ROWS_BLOCK_SIZE;
+    const dim3 block_size(CUDA_SET_ROWS_BLOCK_SIZE);
+    const dim3 grid_size(num_blocks);
+
+
+    const int64_t s01 = nb01/sizeof(src_t);
+    const int64_t s02 = nb02/sizeof(src_t);
+    const int64_t s03 = nb03/sizeof(src_t);
+    const int64_t s10 = nb10/sizeof(int64_t);
+    const int64_t s11 = nb11/sizeof(int64_t);
+    const int64_t s12 = nb12/sizeof(int64_t);
+    const int64_t s1  = nb1/sizeof(dst_t);
+    const int64_t s2  = nb2/sizeof(dst_t);
+    const int64_t s3  = nb3/sizeof(dst_t);
+
+    if (ne_total > 0) {
+        k_set_rows<<<grid_size, block_size, 0, stream>>>(
+            src0_d, src1_d, dst_d,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            s01, s02, s03,
+            s10, s11, s12,
+            s1, s2, s3);
+    }
+}
+
+
+void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_I64);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const float * src0_d   = (const float *)src0->data;
+    const int64_t * src1_d = (const int64_t *)src1->data;
+
+    cudaStream_t stream = ctx.stream();
+
+
+
+    if (dst->type == GGML_TYPE_F32) {
+        set_rows_cuda(
+            src0_d, src1_d, (float*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_F16) {
+        set_rows_cuda(
+            src0_d, src1_d, (half*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_BF16) {
+        set_rows_cuda(
+            src0_d, src1_d, (nv_bfloat16*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q4_0) {
+        set_rows_cuda_quant<block_q4_0, QK4_0, quantize_f32_q4_0_block>(
+            src0_d, src1_d, (block_q4_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q4_1) {
+        set_rows_cuda_quant<block_q4_1, QK4_1, quantize_f32_q4_1_block>(
+            src0_d, src1_d, (block_q4_1*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q5_0) {
+        set_rows_cuda_quant<block_q5_0, QK5_0, quantize_f32_q5_0_block>(
+            src0_d, src1_d, (block_q5_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q5_1) {
+        set_rows_cuda_quant<block_q5_1, QK5_1, quantize_f32_q5_1_block>(
+            src0_d, src1_d, (block_q5_1*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_Q8_0) {
+        set_rows_cuda_quant<block_q8_0, QK8_0, quantize_f32_q8_0_block>(
+            src0_d, src1_d, (block_q8_0*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else if (dst->type == GGML_TYPE_IQ4_NL) {
+        set_rows_cuda_quant<block_iq4_nl, QK4_NL, quantize_f32_iq4_nl_block>(
+            src0_d, src1_d, (block_iq4_nl*)dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
+    } else {
+        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
+    }
+}

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

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

ggml/src/ggml-cuda/softmax.cu

@@ -2,6 +2,7 @@
 #include "ggml.h"
 #include "softmax.cuh"
 #include <cstdint>
+#include <utility>
 
 template <typename T>
 static __device__ __forceinline__ float t2f32(T val) {
@@ -13,6 +14,29 @@ __device__ float __forceinline__ t2f32<half>(half val) {
     return __half2float(val);
 }
 
+struct soft_max_params {
+
+    int64_t nheads;
+    uint32_t n_head_log2;
+    int64_t ncols;
+    int64_t nrows_x;
+    int64_t nrows_y;
+    int64_t ne00;
+    int64_t ne01;
+    int64_t ne02;
+    int64_t ne03;
+    int64_t nb11;
+    int64_t nb12;
+    int64_t nb13;
+
+    int64_t ne12;
+    int64_t ne13;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+};
+
 // When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
 // As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
 #ifdef __clang__
@@ -21,16 +45,24 @@ __device__ float __forceinline__ t2f32<half>(half val) {
 #endif // __clang__
 template <bool use_shared, int ncols_template, int block_size_template, typename T>
 static __global__ void soft_max_f32(
-        const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
-        const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
-    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+        const float * x, const T * mask, float * dst, const soft_max_params p) {
+    const int ncols = ncols_template == 0 ? p.ncols : ncols_template;
 
     const int tid  = threadIdx.x;
-    const int rowx = blockIdx.x;
-    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
+
+    const int64_t i03 = blockIdx.z;
+    const int64_t i02 = blockIdx.y;
+    const int64_t i01 = blockIdx.x;
+
+    //TODO: noncontigous inputs/outputs
+    const int rowx = blockIdx.x + blockIdx.y * gridDim.x + blockIdx.z * gridDim.x * gridDim.y;
+
+    const int64_t i11 = i01;
+    const int64_t i12 = i02 % p.ne12;
+    const int64_t i13 = i03 % p.ne13;
 
     x    += int64_t(rowx)*ncols;
-    mask += int64_t(rowy)*ncols * (mask != nullptr);
+    mask += (i11*p.nb11 + i12*p.nb12 + i13*p.nb13) / sizeof(T) * (mask != nullptr);
     dst  += int64_t(rowx)*ncols;
 
     const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
@@ -38,7 +70,7 @@ static __global__ void soft_max_f32(
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(p.max_bias, i02, p.n_head_log2, p.m0, p.m1);
 
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
@@ -55,7 +87,7 @@ static __global__ void soft_max_f32(
             break;
         }
 
-        const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
+        const float val = x[col]*p.scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
 
         vals[col] = val;
         max_val = max(max_val, val);
@@ -150,64 +182,58 @@ static __global__ void soft_max_back_f32(
     }
 }
 
+template<int... Ns, typename T>
+static void launch_soft_max_kernels(const float * x, const T * mask, float * dst,
+                             const soft_max_params & p, cudaStream_t stream, dim3 block_dims, dim3 block_nums, size_t nbytes_shared)
+{
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    auto launch_kernel = [=](auto I) -> bool {
+        constexpr int ncols = decltype(I)::value;
+        constexpr int block = (ncols > 1024 ? 1024 : ncols);
+
+        if (p.ncols == ncols) {
+            CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, ncols, block, T>), smpbo);
+            soft_max_f32<true, ncols, block><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, mask, dst, p);
+            return true;
+        }
+        return false;
+    };
+
+    // unary fold over launch_kernel
+    if ((launch_kernel(std::integral_constant<int, Ns>{}) || ...)) {
+        return;
+    }
+
+    //default case
+    CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, 0, 0, T>), smpbo);
+    soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, dst, p);
+}
+
+
 template<typename T>
-static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
     int nth = WARP_SIZE;
+    const int64_t ncols_x = params.ncols;
+
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
-    const dim3 block_nums(nrows_x, 1, 1);
+    const dim3 block_nums(params.ne01, params.ne02, params.ne03);
     const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
-    const uint32_t n_head      = nrows_x/nrows_y;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
-    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
-    if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
-        switch (ncols_x) {
-            case 32:
-                soft_max_f32<true,   32,   32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 64:
-                soft_max_f32<true,   64,   64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 128:
-                soft_max_f32<true,  128,  128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 256:
-                soft_max_f32<true,  256,  256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 512:
-                soft_max_f32<true,  512,  512><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            default:
-                soft_max_f32<true,    0,    0><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-        }
+
+    if (nbytes_shared <= smpbo) {
+        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, dst, params, stream, block_dims, block_nums, nbytes_shared);
     } else {
         const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
     }
 }
 
@@ -235,10 +261,11 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
 
-    const int64_t ne00    = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
     const int64_t nrows_y = src0->ne[1];
 
+    const int64_t ne00 = src0->ne[0];
+
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
@@ -247,10 +274,44 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
+
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
+
+    const uint32_t n_head      = src0->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+
+    soft_max_params params = {};
+    params.nheads = src0->ne[2];
+    params.n_head_log2 = n_head_log2;
+    params.ncols = ne00;
+    params.nrows_x = nrows_x;
+    params.nrows_y = nrows_y;
+    params.ne00 = src0->ne[0];
+    params.ne01 = src0->ne[1];
+    params.ne02 = src0->ne[2];
+    params.ne03 = src0->ne[3];
+    params.nb11 = nb11;
+    params.nb12 = nb12;
+    params.nb13 = nb13;
+    params.ne12 = ne12;
+    params.ne13 = ne13;
+    params.scale = scale;
+    params.max_bias = max_bias;
+    params.m0 = m0;
+    params.m1 = m1;
+
     if (use_f16) {
-        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, params, stream);
     } else {
-        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, params, stream);
     }
 }
 

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

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

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

@@ -4,16 +4,15 @@ template <size_t splitD, size_t N>
 __global__ void __launch_bounds__(splitD, 2)
     ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
                  const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
-                 const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2,
-                 const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
-                 const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
-                 float * __restrict__ dst, const int64_t L) {
-    GGML_UNUSED(src1_nb0);
-    GGML_UNUSED(src2_nb0);
+                 const int32_t * __restrict__ src6, float * __restrict__ dst,
+                 const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+                 const int src2_nb1, const int src2_nb2, const int src3_nb1,
+                 const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+                 const int64_t s_off, const int64_t d_inner, const int64_t L) {
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    const int bidx = blockIdx.x;  // split along B
-    const int bidy = blockIdx.y;  // split along D
+    const int bidx = blockIdx.x;  // split along B (sequences)
+    const int bidy = blockIdx.y;  // split along D (d_inner)
     const int tid  = threadIdx.x;
     const int wid  = tid / 32;
     const int wtid = tid % 32;
@@ -24,23 +23,23 @@ __global__ void __launch_bounds__(splitD, 2)
     float *                 smem_A     = smem;
     float *                 smem_s0    = smem_A + splitD * stride_sA;
 
-    const float * s0_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
-    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
+    const float * s0_block = (const float *) ((const char *) src0 + src6[bidx] * src0_nb3 + bidy * splitD * src0_nb2);
+    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb3) + bidy * splitD * sizeof(float));
     const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float));
     const float * A_block  = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1);
-    const float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb2));
-    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb2));
-    float *       y_block  = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float));
-    float *       s_block  = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1);
+    const float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb3));
+    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb3));
+    float *       y_block  = (float *) ((char *) dst + (bidx * d_inner * L * sizeof(float)) + bidy * splitD * sizeof(float));
+    float *       s_block  = (float *) ((char *) dst + s_off + bidx * src0_nb3 + bidy * splitD * src0_nb2);
 
-    const int stride_s0 = src0_nb1 / sizeof(float);
-    const int stride_x  = src1_nb1 / sizeof(float);
+    const int stride_s0 = src0_nb2 / sizeof(float);
+    const int stride_x  = src1_nb2 / sizeof(float);
     const int stride_dt = src2_nb1 / sizeof(float);
     const int stride_A  = src3_nb1 / sizeof(float);
-    const int stride_B  = src4_nb1 / sizeof(float);
-    const int stride_C  = src5_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
     const int stride_s  = stride_s0;
-    const int stride_y  = stride_x;
+    const int stride_y  = d_inner;
 
     // can N not be 16? for example 32?
     if (N == 16) {
@@ -84,24 +83,167 @@ __global__ void __launch_bounds__(splitD, 2)
     }
 }
 
+// assumes as many threads as d_state
+template <int splitH, int d_state>
+__global__ void __launch_bounds__(d_state, 1)
+    ssm_scan_f32_group(
+        const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2,
+        const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5,
+        const int32_t * __restrict__ src6, float * __restrict__ dst,
+        const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+        const int src2_nb1, const int src2_nb2, const int src3_nb1,
+        const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+        const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok) {
+
+    const int head_idx = (blockIdx.x * splitH) / d_head;
+    const int head_off = ((blockIdx.x * splitH) % d_head) * sizeof(float);
+    const int seq_idx = blockIdx.y;
+
+    const int group_off = (head_idx & (n_group - 1)) * d_state * sizeof(float);
+
+    const float * s0_block = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+    const float * x_block  = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + blockIdx.x * splitH * sizeof(float));
+    const float * dt_block = (const float *) ((const char *) src2 + (seq_idx * src2_nb2) + head_idx * sizeof(float));
+    const float * A_block  = (const float *) ((const char *) src3 + head_idx * src3_nb1);
+    const float * B_block  = (const float *) ((const char *) src4 + (seq_idx * src4_nb3) + (group_off));
+    const float * C_block  = (const float *) ((const char *) src5 + (seq_idx * src5_nb3) + (group_off));
+    float *       y_block  = dst + (seq_idx * n_tok * n_head * d_head) + blockIdx.x * splitH;
+    float *       s_block  = (float *) ((char *) dst + s_off + seq_idx * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+
+    // strides across n_seq_tokens
+    const int stride_x  = src1_nb2 / sizeof(float);
+    const int stride_dt = src2_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
+    const int stride_y  = n_head * d_head;
+
+    float state[splitH];
+    // for the parallel accumulation
+    __shared__ float stateC[splitH * d_state];
+
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        state[j] = s0_block[j * d_state + threadIdx.x];
+    }
+
+    for (int64_t i = 0; i < n_tok; i++) {
+        // TODO: only calculate dA and dt_soft_plus once per head instead of every splitH head elements
+        // TODO: only calculate B and C once per head group
+        // NOTE: dt_soft_plus, dA and x_dt have the same value across threads here.
+        float dt_soft_plus = dt_block[i * stride_dt];
+        if (dt_soft_plus <= 20.0f) {
+            dt_soft_plus = log1pf(expf(dt_soft_plus));
+        }
+        const float dA = expf(dt_soft_plus * A_block[0]);
+        const float B = B_block[i * stride_B + threadIdx.x];
+        const float C = C_block[i * stride_C + threadIdx.x];
+
+        // across d_head
+#pragma unroll
+        for (int j = 0; j < splitH; j++) {
+            const float x_dt = x_block[i * stride_x + j] * dt_soft_plus;
+
+            state[j] = (state[j] * dA) + (B * x_dt);
+
+            stateC[j * d_state + threadIdx.x] = state[j] * C;
+        }
+
+        __syncthreads();
+
+        // parallel accumulation for stateC
+        // TODO: simplify
+        {
+            static_assert((d_state & -d_state) == d_state, "the state size has to be a power of 2");
+            static_assert((splitH & -splitH) == splitH, "splitH has to be a power of 2");
+
+            // reduce until w matches the warp size
+            // TODO: does this work even when the physical warp size is 64?
+#pragma unroll
+            for (int w = d_state; w > WARP_SIZE; w >>= 1) {
+                // (assuming there are d_state threads)
+#pragma unroll
+                for (int j = 0; j < ((w >> 1) * splitH + d_state - 1) / d_state; j++) {
+                    // TODO: check for bank conflicts
+                    const int k = (threadIdx.x % (w >> 1)) + (d_state * (threadIdx.x / (w >> 1))) + j * d_state * (d_state / (w >> 1));
+                    stateC[k] += stateC[k + (w >> 1)];
+
+                }
+                __syncthreads();
+            }
+
+            static_assert(splitH >= d_state / WARP_SIZE);
+
+#pragma unroll
+            for (int j = 0; j < splitH / (d_state / WARP_SIZE); j++) {
+                float y = stateC[(threadIdx.x % WARP_SIZE) + d_state * (threadIdx.x / WARP_SIZE) + j * d_state * (d_state / WARP_SIZE)];
+                y = warp_reduce_sum(y);
+
+                // store the above accumulations
+                if (threadIdx.x % WARP_SIZE == 0) {
+                    const int k = threadIdx.x / WARP_SIZE + j * (d_state / WARP_SIZE);
+                    y_block[i * stride_y + k] = y;
+                }
+            }
+        }
+    }
+
+    // write back the state
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        s_block[j * d_state + threadIdx.x] = state[j];
+    }
+}
+
 static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3,
-                              const float * src4, const float * src5, const int src0_nb1, const int src0_nb2,
-                              const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3,
-                              const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1,
-                              const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2,
-                              float * dst, const int64_t N, const int64_t D, const int64_t L, const int64_t B,
+                              const float * src4, const float * src5, const int32_t * src6, float * dst,
+                              const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3, const int src2_nb1,
+                              const int src2_nb2, const int src3_nb1, const int src4_nb2, const int src4_nb3, const int src5_nb2,
+                              const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
+                              const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
                               cudaStream_t stream) {
-    const int threads = 128;
-    // todo: consider D cannot be divided,does this situation exist?
-    GGML_ASSERT(D % threads == 0);
-    const dim3 blocks(B, (D + threads - 1) / threads, 1);
-    const int  smem_size = (threads * (N + 1) * 2) * sizeof(float);
-    if (N == 16) {
-        ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
-            src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0,
-            src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, L);
+    // NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
+    if (src3_nb1 == sizeof(float)) {
+        // Mamba-2
+        if (d_state == 128) {
+            const int threads = 128;
+            GGML_ASSERT(d_state % threads == 0);
+            // NOTE: can be any power of two between 4 and 64
+            const int splitH = 16;
+            GGML_ASSERT(head_dim % splitH == 0);
+            const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
+            ssm_scan_f32_group<16, 128><<<blocks, threads, 0, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
+        } else if (d_state == 256) { // Falcon-H1
+            const int threads = 256;
+            // NOTE: can be any power of two between 8 and 64
+            const int splitH = 16;
+            GGML_ASSERT(head_dim % splitH == 0);
+            const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
+            ssm_scan_f32_group<16, 256><<<blocks, threads, 0, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
+        } else {
+            GGML_ABORT("doesn't support d_state!=(128 or 256).");
+        }
     } else {
-        GGML_ABORT("doesn't support N!=16.");
+        const int threads = 128;
+        // Mamba-1
+        GGML_ASSERT(n_head % threads == 0);
+        GGML_ASSERT(head_dim == 1);
+        GGML_ASSERT(n_group == 1);
+        const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
+        const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
+        if (d_state == 16) {
+            ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
+                src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+        } else {
+            GGML_ABORT("doesn't support d_state!=16.");
+        }
     }
 }
 
@@ -112,30 +254,25 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const struct ggml_tensor * src3 = dst->src[3];  // A
     const struct ggml_tensor * src4 = dst->src[4];  // B
     const struct ggml_tensor * src5 = dst->src[5];  // C
-
-    //   const int64_t d_state = src0->ne[0];
-    //   const int64_t d_inner = src0->ne[1];
-    //   const int64_t l = src1->ne[1];
-    //   const int64_t b = src0->ne[2];
+    const struct ggml_tensor * src6 = dst->src[6];  // ids
 
     const int64_t nc  = src0->ne[0];  // d_state
-    const int64_t nr  = src0->ne[1];  // d_inner
-    const int64_t n_t = src1->ne[1];  // number of tokens per sequence
-    const int64_t n_s = src0->ne[2];  // number of sequences in the batch
+    const int64_t nr  = src0->ne[1];  // head_dim or 1
+    const int64_t nh  = src1->ne[1];  // n_head
+    const int64_t ng  = src4->ne[1];  // n_group
+    const int64_t n_t = src1->ne[2];  // number of tokens per sequence
+    const int64_t n_s = src1->ne[3];  // number of sequences in the batch
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    const int64_t s_off = ggml_nelements(src1) * sizeof(float);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*n_s == ggml_nelements(dst));
     GGML_ASSERT(src0->nb[0] == sizeof(float));
     GGML_ASSERT(src1->nb[0] == sizeof(float));
     GGML_ASSERT(src2->nb[0] == sizeof(float));
     GGML_ASSERT(src3->nb[0] == sizeof(float));
     GGML_ASSERT(src4->nb[0] == sizeof(float));
     GGML_ASSERT(src5->nb[0] == sizeof(float));
-    // required for the dot product between s and C
-    GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float));
-    // required for per-sequence offsets for states
-    GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float));
-    // required to get correct offset for state destination (i.e. src1->nb[3])
-    GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float));
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
 
     const float * src0_d = (const float *) src0->data;
     const float * src1_d = (const float *) src1->data;
@@ -143,13 +280,16 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float * src3_d = (const float *) src3->data;
     const float * src4_d = (const float *) src4->data;
     const float * src5_d = (const float *) src5->data;
+    const int32_t * src6_d = (const int32_t *) src6->data;
     float *       dst_d  = (float *) dst->data;
     cudaStream_t  stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src6->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
-    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0],
-                      src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1],
-                      src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream);
+    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
+                      src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
+                      src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],
+                      s_off, nc, nr, nh, ng, n_t, n_s, stream);
 }

ggml/src/ggml-cuda/unary.cu

@@ -83,6 +83,10 @@ static __device__ __forceinline__ float op_log(float x) {
     return logf(x);
 }
 
+static __device__ __forceinline__ float op_elu(float x) {
+    return (x > 0.f) ? x : expm1f(x);
+}
+
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -196,6 +200,9 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_log>(ctx, dst);
 }
 
+void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_elu>(ctx, dst);
+}
 /* gated ops */
 
 template <float (*op)(float), typename T>
@@ -285,6 +292,14 @@ void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary_gated<op_silu>(ctx, dst);
 }
 
+void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu_erf>(ctx, dst);
+}
+
+void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu_quick>(ctx, dst);
+}
+
 /* silu_back */
 
 static __device__ __forceinline__ float op_silu_back(float grad, float x) {

ggml/src/ggml-cuda/unary.cuh

@@ -59,8 +59,14 @@ void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/upscale.cu

@@ -22,17 +22,88 @@ static __global__ void upscale_f32(const float * x, float * dst,
     dst[index] = *( (const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00) );
 }
 
+static __global__ void upscale_f32_bilinear(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset) {
+    const int64_t index              = threadIdx.x + blockIdx.x * blockDim.x;
+    const int64_t dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    const int i10_dst = index % ne10_dst;
+    const int i11_dst = (index / ne10_dst) % ne11_dst;
+    const int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    const int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    const int i02_src = (int)(i12_dst / sf2);
+    const int i03_src = (int)(i13_dst / sf3);
+
+    const float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    int y0_src    = (int)floorf(y_src_f);
+    int y1_src    = y0_src + 1;
+
+    y0_src = max(0, min(y0_src, ne01_src - 1));
+    y1_src = max(0, min(y1_src, ne01_src - 1));
+
+    float dy = y_src_f - (float)y0_src;
+    dy       = max(0.0f, min(dy, 1.0f));
+
+    float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    int x0_src    = (int)floorf(x_src_f);
+    int x1_src    = x0_src + 1;
+
+    x0_src = max(0, min(x0_src, ne00_src - 1));
+    x1_src = max(0, min(x1_src, ne00_src - 1));
+
+    float dx = x_src_f - (float)x0_src;
+    dx = max(0.0f, min(dx, 1.0f));
+
+    const float * p_a = (const float *)((const char *)x + (int64_t)x0_src * nb00 + (int64_t)y0_src * nb01 + (int64_t)i02_src * nb02 + (int64_t)i03_src * nb03);
+    const float * p_b = (const float *)((const char *)x + (int64_t)x1_src * nb00 + (int64_t)y0_src * nb01 + (int64_t)i02_src * nb02 + (int64_t)i03_src * nb03);
+    const float * p_c = (const float *)((const char *)x + (int64_t)x0_src * nb00 + (int64_t)y1_src * nb01 + (int64_t)i02_src * nb02 + (int64_t)i03_src * nb03);
+    const float * p_d = (const float *)((const char *)x + (int64_t)x1_src * nb00 + (int64_t)y1_src * nb01 + (int64_t)i02_src * nb02 + (int64_t)i03_src * nb03);
+
+    const float val_a = *p_a;
+    const float val_b = *p_b;
+    const float val_c = *p_c;
+    const float val_d = *p_d;
+
+    float result = val_a * (1.0f - dx) * (1.0f - dy) +
+                   val_b * dx * (1.0f - dy) +
+                   val_c * (1.0f - dx) * dy +
+                   val_d * dx * dy;
+
+    dst[index] = result;
+}
+
 static void upscale_f32_cuda(const float * x, float * dst,
         const int nb00, const int nb01, const int nb02, const int nb03,
         const int ne10, const int ne11, const int ne12, const int ne13,
         const float sf0, const float sf1, const float sf2, const float sf3,
         cudaStream_t stream) {
-    int dst_size = ne10 * ne11 * ne12 * ne13;
-    int num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+    const int64_t dst_size   = ne10 * ne11 * ne12 * ne13;
+    const int64_t num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
 
     upscale_f32<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
 }
 
+static void upscale_f32_bilinear_cuda(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset, cudaStream_t stream) {
+    const int64_t dst_size   = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    const int64_t num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+
+    upscale_f32_bilinear<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+}
+
 void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;
@@ -42,10 +113,25 @@ void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const float sf0 = (float)dst->ne[0]/src0->ne[0];
-    const float sf1 = (float)dst->ne[1]/src0->ne[1];
-    const float sf2 = (float)dst->ne[2]/src0->ne[2];
+    const int mode_flags = dst->op_params[0];
+    const ggml_scale_mode mode = (ggml_scale_mode)(mode_flags & 0xFF);
+
+    float sf0 = (float)dst->ne[0]/src0->ne[0];
+    float sf1 = (float)dst->ne[1]/src0->ne[1];
+    float sf2 = (float)dst->ne[2]/src0->ne[2];
     const float sf3 = (float)dst->ne[3]/src0->ne[3];
 
-    upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        float pixel_offset = 0.5f;
+        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+            sf0          = (float)(dst->ne[0] - 1) / (src0->ne[0] - 1);
+            sf1          = (float)(dst->ne[1] - 1) / (src0->ne[1] - 1);
+            pixel_offset = 0.0f;
+        }
+        upscale_f32_bilinear_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                                 src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+                                 sf0, sf1, sf2, sf3, pixel_offset, stream);
+    }
 }

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

@@ -10,9 +10,6 @@
 #include "rocblas/rocblas.h"
 #endif // __HIP_PLATFORM_AMD__
 
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N HIPBLAS_OP_N
@@ -30,7 +27,6 @@
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
 #define cublasCreate hipblasCreate
 #define cublasDestroy hipblasDestroy
 #define cublasGemmEx hipblasGemmEx
@@ -42,7 +38,6 @@
 #define cublasSgemm hipblasSgemm
 #define cublasStatus_t hipblasStatus_t
 #define cublasOperation_t hipblasOperation_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
 #define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
@@ -144,6 +139,20 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION >= 70000000
+#define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
+#define cublasComputeType_t hipblasComputeType_t
+#define cudaDataType_t hipDataType
+#else
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define cublasComputeType_t hipblasDatatype_t
+#define cudaDataType_t hipblasDatatype_t
+#endif
+
 #define __CUDA_ARCH__ 1300
 
 #if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)

ggml/src/ggml-impl.h

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

ggml/src/ggml-kompute/CMakeLists.txt

@@ -1,166 +0,0 @@
-
-find_package(Vulkan COMPONENTS glslc REQUIRED)
-find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
-
-if (NOT glslc_executable)
-    message(FATAL_ERROR "glslc not found")
-endif()
-
-ggml_add_backend_library(ggml-kompute
-                         ggml-kompute.cpp
-                         ../../include/ggml-kompute.h
-                        )
-
-target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
-target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
-
-add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
-
-function(compile_shader)
-    set(options)
-    set(oneValueArgs)
-    set(multiValueArgs SOURCES)
-    cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-    foreach(source ${compile_shader_SOURCES})
-        get_filename_component(filename ${source} NAME)
-        set(spv_file ${filename}.spv)
-        add_custom_command(
-            OUTPUT ${spv_file}
-            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-            COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            COMMENT "Compiling ${source} to ${spv_file}"
-            )
-
-        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-        set(FILE_NAME "shader${RAW_FILE_NAME}")
-        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-        if(CMAKE_GENERATOR MATCHES "Visual Studio")
-            add_custom_command(
-                OUTPUT ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                DEPENDS ${spv_file} xxd
-                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-                )
-        else()
-            add_custom_command(
-                OUTPUT ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                DEPENDS ${spv_file} xxd
-                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-                )
-        endif()
-    endforeach()
-endfunction()
-
-if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
-    message(STATUS "Kompute found")
-    set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
-    add_subdirectory(kompute)
-
-    # Compile our shaders
-    compile_shader(SOURCES
-        kompute-shaders/op_scale.comp
-        kompute-shaders/op_scale_8.comp
-        kompute-shaders/op_add.comp
-        kompute-shaders/op_addrow.comp
-        kompute-shaders/op_mul.comp
-        kompute-shaders/op_silu.comp
-        kompute-shaders/op_relu.comp
-        kompute-shaders/op_gelu.comp
-        kompute-shaders/op_softmax.comp
-        kompute-shaders/op_norm.comp
-        kompute-shaders/op_rmsnorm.comp
-        kompute-shaders/op_diagmask.comp
-        kompute-shaders/op_mul_mat_mat_f32.comp
-        kompute-shaders/op_mul_mat_f16.comp
-        kompute-shaders/op_mul_mat_q8_0.comp
-        kompute-shaders/op_mul_mat_q4_0.comp
-        kompute-shaders/op_mul_mat_q4_1.comp
-        kompute-shaders/op_mul_mat_q4_k.comp
-        kompute-shaders/op_mul_mat_q6_k.comp
-        kompute-shaders/op_getrows_f32.comp
-        kompute-shaders/op_getrows_f16.comp
-        kompute-shaders/op_getrows_q4_0.comp
-        kompute-shaders/op_getrows_q4_1.comp
-        kompute-shaders/op_getrows_q6_k.comp
-        kompute-shaders/op_rope_norm_f16.comp
-        kompute-shaders/op_rope_norm_f32.comp
-        kompute-shaders/op_rope_neox_f16.comp
-        kompute-shaders/op_rope_neox_f32.comp
-        kompute-shaders/op_cpy_f16_f16.comp
-        kompute-shaders/op_cpy_f16_f32.comp
-        kompute-shaders/op_cpy_f32_f16.comp
-        kompute-shaders/op_cpy_f32_f32.comp
-    )
-
-    # Create a custom target for our generated shaders
-    add_custom_target(generated_shaders DEPENDS
-        shaderop_scale.h
-        shaderop_scale_8.h
-        shaderop_add.h
-        shaderop_addrow.h
-        shaderop_mul.h
-        shaderop_silu.h
-        shaderop_relu.h
-        shaderop_gelu.h
-        shaderop_softmax.h
-        shaderop_norm.h
-        shaderop_rmsnorm.h
-        shaderop_diagmask.h
-        shaderop_mul_mat_mat_f32.h
-        shaderop_mul_mat_f16.h
-        shaderop_mul_mat_q8_0.h
-        shaderop_mul_mat_q4_0.h
-        shaderop_mul_mat_q4_1.h
-        shaderop_mul_mat_q4_k.h
-        shaderop_mul_mat_q6_k.h
-        shaderop_getrows_f32.h
-        shaderop_getrows_f16.h
-        shaderop_getrows_q4_0.h
-        shaderop_getrows_q4_1.h
-        shaderop_getrows_q6_k.h
-        shaderop_rope_norm_f16.h
-        shaderop_rope_norm_f32.h
-        shaderop_rope_neox_f16.h
-        shaderop_rope_neox_f32.h
-        shaderop_cpy_f16_f16.h
-        shaderop_cpy_f16_f32.h
-        shaderop_cpy_f32_f16.h
-        shaderop_cpy_f32_f32.h
-    )
-
-    # Create a custom command that depends on the generated_shaders
-    add_custom_command(
-        OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-        COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-        DEPENDS generated_shaders
-        COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
-    )
-
-    # Add the stamp to the main sources to ensure dependency tracking
-    target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-else()
-    message(WARNING "Kompute not found")
-endif()

ggml/src/ggml-kompute/kompute-shaders/common.comp

@@ -1,112 +0,0 @@
-#extension GL_EXT_shader_16bit_storage: require
-#extension GL_EXT_shader_8bit_storage: require
-#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int64: require
-#extension GL_EXT_control_flow_attributes: enable
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-#define QK4_0 32
-#define QK4_1 32
-
-#define GELU_COEF_A 0.044715
-#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
-#define TWOPI_F 6.283185307179586f
-
-#define QK_K 256
-#define K_SCALE_SIZE 12
-
-#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
-#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
-#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
-#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
-
-#define sizeof_block_q4_0 0x12
-struct block_q4_0 {
-    float16_t d;
-    uint8_t qs[QK4_0 / 2];
-};
-mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float md = -8.f * xb.d;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
-        reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
-    }
-    return reg;
-}
-
-#define sizeof_block_q4_1 0x14
-struct block_q4_1 {
-    float16_t d;
-    float16_t m;
-    uint8_t qs[QK4_1 / 2];
-};
-mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float  m = xb.m;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
-        reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
-    }
-    return reg;
-}
-
-#define sizeof_block_q4_k 144
-struct block_q4_k {
-    float16_t d;
-    float16_t dmin;
-    uint8_t scales[K_SCALE_SIZE];
-    uint8_t qs[QK_K/2];
-};
-
-#define sizeof_block_q6_k 210
-struct block_q6_k {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    float16_t d;             // super-block scale
-};
-mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
-    const float16_t d_all = xb.d;
-
-    const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
-    const uint qhIndex = 32*(il/8) + 16*(il&1);
-    float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
-    il = (il/2) & 3;
-
-    const uint16_t  kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
-    const uint16_t  kmask2 = il>1 ? uint8_t(0xF0)             : uint8_t(0x0F);
-    const float16_t coef   = il>1 ? float16_t(1.f/16.f)       : float16_t(1.f);
-    const float16_t ml = float16_t(d_all * sc * 32.f);
-    const float16_t dl = float16_t(d_all * sc * coef);
-    mat4 reg;
-    for (int i = 0; i < 16; ++i) {
-        const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
-                                        : ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
-        reg[i/4][i%4] = dl * q - ml;
-    }
-    return reg;
-}
-
-
-#define QK8_0 32
-// struct block_q8_0 {
-//     float16_t d;         // delta
-//     int8_t    qs[QK8_0]; // quants
-// };
-#define sizeof_block_q8_0 34

ggml/src/ggml-kompute/kompute-shaders/op_add.comp

@@ -1,58 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-  //int offs; // TODO: needed for GGML_OP_ACC, see metal code
-} pcs;
-
-// general-purpose kernel for addition of two tensors
-// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
-// cons: not very efficient
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    int offs = 0; // TMP (see above)
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11       ) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1  + offs) / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp

@@ -1,25 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    uint row;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp

@@ -1,30 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint n_past;
-    int ne00;
-    int ne01;
-} pcs;
-
-void main() {
-    const uint i02 = gl_WorkGroupID.z;
-    const uint i01 = gl_WorkGroupID.y;
-    const uint i00 = gl_WorkGroupID.x;
-
-    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
-
-    if (i00 > pcs.n_past + i01) {
-        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
-    } else {
-        out_[index + pcs.outOff] = in_[index + pcs.inOff];
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp

@@ -1,22 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 8;
-
-    for (uint x = 0; x < 8; x++) {
-        const uint i = baseIndex + x;
-        const float y = in_[i + pcs.inOff];
-        out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp

@@ -1,17 +0,0 @@
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    int z = 0;
-    for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
-        const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
-        const mat4 result = dequantize_block(inIndex, ind%NL);
-        for (uint j = 0; j < 4; ++j) {
-            for (uint k = 0; k < 4; ++k) {
-                const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
-                out_[outIndex] = result[j][k];
-                ++z;
-            }
-        }
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp

@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}

ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp

@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}

ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp

@@ -1,38 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_0 get_unaligned_block_q4_0(uint index) {
-    block_q4_0 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
-        fres.qs[it] = inA[index+2+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_0 block = get_unaligned_block_q4_0(index);
-    return dequantize_q4_0(block, il);
-}
-
-#include "op_getrows.comp"

ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp

@@ -1,39 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_1 get_unaligned_block_q4_1(uint index) {
-    block_q4_1 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    fres.m = u8BufToFloat16(inA, index+2);
-    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
-        fres.qs[it] = inA[index+4+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_1 block = get_unaligned_block_q4_1(index);
-    return dequantize_q4_1(block, il);
-}
-
-#include "op_getrows.comp"

ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp

@@ -1,44 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 16
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q6_k get_unaligned_block_q6_k(uint index) {
-    block_q6_k fres;
-    [[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
-        fres.ql[it] = inA[index + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
-        fres.qh[it] = inA[index + QK_K/2 + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
-        fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
-    }
-    fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q6_k block = get_unaligned_block_q6_k(index);
-    return dequantize_q6_k(block, il);
-}
-
-#include "op_getrows.comp"

ggml/src/ggml-kompute/kompute-shaders/op_mul.comp

@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1)  / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp

@@ -1,69 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-
-layout(local_size_x_id = 0) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    uint nb10;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    int ne0;
-    int ne1;
-    uint r2;
-    uint r3;
-} pcs;
-
-#define N_F16_F32 4
-
-void main() {
-    const uint r0 = gl_WorkGroupID.x;
-    const uint rb = gl_WorkGroupID.y*N_F16_F32;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb03;
-
-    const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
-
-    for (uint row = 0; row < N_F16_F32; ++row) {
-        uint r1 = rb + row;
-        if (r1 >= pcs.ne11) {
-            break;
-        }
-
-        const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-        float sumf = 0;
-        for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-            sumf += float(inA[x+i]) * float(inB[y+i]);
-        }
-
-        const float all_sum = subgroupAdd(sumf);
-        if (subgroupElect()) {
-            out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
-        }
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp

@@ -1,51 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-// device subgroup size
-layout (local_size_x_id = 0) in;
-
-layout(binding = 0) readonly buffer tensorInA { float inA[]; };
-layout(binding = 1) readonly buffer tensorInB { float inB[]; };
-layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout(push_constant) uniform parameter {
-  uint inAOff;
-  uint inBOff;
-  uint outOff;
-  int ne00;
-  int ne01;
-  int ne02;
-  int ne11;
-  int ne12;
-  uint nb01;
-  uint nb02;
-  uint nb11;
-  uint nb12;
-  uint nb1;
-  uint nb2;
-}
-pcs;
-
-
-void main() {
-  uvec3 gid = gl_WorkGroupID;
-
-  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
-  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
-
-  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
-  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
-  float sum = 0.0f;
-  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-      sum += float(inA[x+i]) * float(inB[y+i]);
-  }
-
-  const float all_sum = subgroupAdd(sum);
-  if (subgroupElect()) {
-    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
-  }
-}

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp

@@ -1,33 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_0
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_0 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (sumy * -8.f + acc[0] + acc[1]);
-}
-
-#include "op_mul_mv_q_n.comp"

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp

@@ -1,35 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_1
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_1 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float m = float(u8BufToFloat16(inA, index+2));
-
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (acc[0] + acc[1]) + sumy * m;
-}
-
-#include "op_mul_mv_q_n.comp"

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp

@@ -1,140 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define N_DST 4
-#define SIZE_OF_BLOCK sizeof_block_q4_k
-
-layout(local_size_x = 4) in;
-layout(local_size_y = 8) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int ne02;
-    int ne12;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
-
-void main() {
-    const uint16_t kmask1 = uint16_t(0x3f3f);
-    const uint16_t kmask2 = uint16_t(0x0f0f);
-    const uint16_t kmask3 = uint16_t(0xc0c0);
-
-    const uint ix = gl_SubgroupInvocationID/8;  // 0...3
-    const uint it = gl_SubgroupInvocationID%8;  // 0...7
-    const uint iq = it/4;     // 0 or 1
-    const uint ir = it%4;     // 0...3
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = r0 * N_DST;
-    const uint ib_row = first_row * nb;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-    const uint offset1 =        r1*pcs.nb11 + (i12       )*pcs.nb12 + (i13       )*pcs.nb13;
-
-    const uint xblk = offset0 + pcs.inAOff;
-    const uint y = (offset1 / 4) + pcs.inBOff;
-
-    float yl[16];
-    float yh[16];
-    float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
-    float all_sum = 0.f;
-
-    uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
-
-    for (uint ib = ix; ib < nb; ib += 4) {
-        const uint blk_idx = ib + xblk;
-
-        float sumy[4] = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+0] = inB[y4+i+  0]; sumy[0] += yl[i+0];
-            yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
-            yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
-            yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
-        }
-
-        for (int row = 0; row < N_DST; row++) {
-            uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK);
-
-            uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
-            uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
-            uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
-            uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
-            uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
-
-            uint16_t sc16[4];
-            sc16[0] = sc_0 & kmask1;
-            sc16[1] = sc_2 & kmask1;
-            sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
-            sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
-
-            float acc1[4] = {0.f, 0.f, 0.f, 0.f};
-            float acc2[4] = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
-                uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
-                acc1[0] += yl[i+0] * (q1 & 0x000F);
-                acc1[1] += yl[i+1] * (q1 & 0x0F00);
-                acc1[2] += yl[i+8] * (q1 & 0x00F0);
-                acc1[3] += yl[i+9] * (q1 & 0xF000);
-                acc2[0] += yh[i+0] * (q2 & 0x000F);
-                acc2[1] += yh[i+1] * (q2 & 0x0F00);
-                acc2[2] += yh[i+8] * (q2 & 0x00F0);
-                acc2[3] += yh[i+9] * (q2 & 0xF000);
-            }
-
-            uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
-            uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
-            uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
-            uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
-            uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
-            uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
-            uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
-            uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
-
-            float dall = float(inA[blk_idx + row_idx].d);
-            float dmin = float(inA[blk_idx + row_idx].dmin);
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
-                               (acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
-                               (acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
-                               (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
-                dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
-        }
-
-        y4 += 4 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = subgroupAdd(sumf[row]);
-        if (subgroupElect()) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
-        }
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp

@@ -1,106 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x_id = 0) in;
-layout(local_size_y_id = 1) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int ne02;
-    int ne12;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
-
-void main() {
-    const uint8_t kmask1 = uint8_t(0x03);
-    const uint8_t kmask2 = uint8_t(0x0C);
-    const uint8_t kmask3 = uint8_t(0x30);
-    const uint8_t kmask4 = uint8_t(0xC0);
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint x = row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-    const uint yy = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-    float sumf = 0;
-
-    // bits of invocation ID for gl_SubgroupSize=32:
-    //  x   x   x   x   x
-    //  4   3   2   1   0
-    // (     tid     ) ix
-    //  ip (   il    )
-
-    const uint block_stride = gl_SubgroupSize / 16;         // number of blocks each subgroup processes
-    const uint tid  = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
-    const uint ix   = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
-    const uint ip   = tid/8;        // first or second half of block (0 or 1)
-    const uint il   = tid%8;        // each half has 8 parts, one per scale
-    const uint n    = 4;            // 4 scales at a time (and 4 sums)
-    const uint l0   = n*il;         // offset into half-block, 0..28
-    const uint is   = 8*ip + l0/16; // 0, 1, 8, 9
-
-    const uint y_offset = 128*ip + l0;
-    const uint q_offset_l = 64*ip + l0;
-    const uint q_offset_h = 32*ip + l0;
-
-    for (uint i = ix; i < nb; i += block_stride) {
-
-        const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
-
-        const uint qlIndex = q_offset_l;
-        const uint q2Index = qlIndex + QK_K/8;
-        const uint qhIndex = q_offset_h;
-        const uint y = yy + i * QK_K + y_offset;
-
-        float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
-        for (uint l = 0; l < n; ++l) {
-            const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
-            const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
-            const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
-
-            sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
-            sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
-            sums[2] += inB[y+l+64] * (int8_t((currentQ1  >> 4) | ((currentQh & kmask3) << 0)) - 32);
-            sums[3] += inB[y+l+96] * (int8_t((currentQ2  >> 4) | ((currentQh & kmask4) >> 2)) - 32);
-        }
-
-        float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
-        sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
-    }
-
-    const float tot = subgroupAdd(sumf);
-    if (subgroupElect()) {
-        out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
-    }
-}

ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp

@@ -1,73 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#include "op_mul_mv_q_n_pre.comp"
-
-#define SIZE_OF_D 2
-
-#define N_DST 4 // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
-#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
-
-#define NB_Q8_0 8
-
-void main() {
-    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const int nr  = N_DST;
-    const int nsg = N_SIMDGROUP;
-    const int nw  = N_SIMDWIDTH;
-
-    const int nb = pcs.ne00/QK8_0;
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
-
-    const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
-    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
-
-    float yl[NB_Q8_0];
-    float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
-
-    const uint ix = gl_SubgroupInvocationID.x/4;
-    const uint il = gl_SubgroupInvocationID.x%4;
-
-    uint yb = y + ix * QK8_0 + NB_Q8_0*il;
-
-    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
-    for (uint ib = ix; ib < nb; ib += nw/4) {
-        for (int i = 0; i < NB_Q8_0; ++i) {
-            yl[i] = inB[yb + i];
-        }
-
-        for (int row = 0; row < nr; row++) {
-            const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
-            float sumq = 0.f;
-            for (int iq = 0; iq < NB_Q8_0; ++iq) {
-                const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
-                sumq += qs_iq * yl[iq];
-            }
-            const float16_t d = u8BufToFloat16(inA, x + block_offset);
-            sumf[row] += sumq*d;
-        }
-
-        yb += NB_Q8_0 * nw;
-    }
-
-    for (int row = 0; row < nr; ++row) {
-        const float tot = subgroupAdd(sumf[row]);
-        if (subgroupElect() && first_row + row < pcs.ne01) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
-        }
-    }
-}