sync : ggml

ggml-ci

.devops/nix/package.nix

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

.devops/tools.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 
 # Read the first argument into a variable

.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]
+        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]
+        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:
@@ -93,3 +87,8 @@ Ascend NPU:
             - ggml/include/ggml-cann.h
             - ggml/src/ggml-cann/**
             - docs/backend/CANN.md
+OpenCL:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-opencl.h
+            - ggml/src/ggml-opencl/**

.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

@@ -84,7 +84,8 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_METAL_EMBED_LIBRARY=OFF \
+            -DGGML_METAL_SHADER_DEBUG=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -134,6 +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 $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
   ubuntu-cpu-cmake:
     strategy:
       matrix:
@@ -341,6 +405,72 @@ jobs:
           cd build
           export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
+          ctest -L main --verbose --timeout 4200
+
+  ubuntu-22-cmake-webgpu:
+    runs-on: ubuntu-22.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: hendrikmuhs/ccache-action@v1.2.16
+        with:
+          key: ubuntu-22-cmake-webgpu
+          evict-old-files: 1d
+
+      - name: Vulkan SDK Dependencies
+        id: vulkan-depends
+        run: |
+          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+          sudo apt-get update -y
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
+          ARTIFACTS_JSON=$(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:
@@ -664,7 +794,7 @@ jobs:
           ./build-xcframework.sh
 
   windows-msys2:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     strategy:
       fail-fast: false
@@ -714,7 +844,7 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
@@ -725,17 +855,20 @@ jobs:
       matrix:
         include:
           - build: 'cpu-x64 (static)'
+            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 -DBUILD_SHARED_LIBS=OFF'
           - build: 'openblas-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_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
+            arch: 'x64'
             defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - 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'
-         #   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
@@ -749,12 +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' }}
@@ -770,7 +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
@@ -805,6 +932,8 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -825,7 +954,7 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
+        if: ${{ matrix.arch == 'x64' }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -930,7 +1059,7 @@ jobs:
           cmake --build build --config Release
 
   windows-latest-cmake-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -964,7 +1093,7 @@ jobs:
 
   windows-latest-cmake-hip:
     if: ${{ github.event.inputs.create_release != 'true' }}
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     steps:
       - name: Clone

.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 \
@@ -235,7 +241,7 @@ jobs:
           name: llama-bin-ubuntu-vulkan-x64.zip
 
   windows-cpu:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     strategy:
       matrix:
@@ -271,7 +277,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
             -DGGML_NATIVE=OFF ^
@@ -288,7 +294,7 @@ jobs:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -298,7 +304,7 @@ jobs:
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
   windows:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
@@ -448,7 +454,7 @@ jobs:
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -520,7 +526,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-hip:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     strategy:
       matrix:

.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

build-xcframework.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # Options
 IOS_MIN_OS_VERSION=16.4

ci/run.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # sample usage:
 #
@@ -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"),
@@ -2794,6 +2809,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ssl_file_cert = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
+    add_opt(common_arg(
+        {"--chat-template-kwargs"}, "STRING",
+        string_format("sets additional params for the json template parser"),
+        [](common_params & params, const std::string &  value) {
+            auto parsed = json::parse(value);
+            for (const auto & item : parsed.items()) {
+                params.default_template_kwargs[item.key()] = item.value().dump();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
     add_opt(common_arg(
         {"-to", "--timeout"}, "N",
         string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),
@@ -3406,5 +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/chat.cpp

@@ -17,6 +17,8 @@
 #include <string>
 #include <vector>
 
+using json = nlohmann::ordered_json;
+
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto time = std::chrono::system_clock::to_time_t(now);
     auto local_time = *std::localtime(&time);
@@ -140,6 +142,7 @@ struct templates_params {
     bool add_generation_prompt = true;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    json extra_context;
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -720,16 +723,23 @@ static void foreach_function(const json & tools, const std::function<void(const
 
 static std::string apply(
     const common_chat_template & tmpl,
-    const nlohmann::ordered_json & messages,
-    const nlohmann::ordered_json & tools,
-    bool add_generation_prompt,
-    const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
+    const struct templates_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt)
 {
     minja::chat_template_inputs tmpl_inputs;
-    tmpl_inputs.messages = messages;
-    tmpl_inputs.tools = tools;
-    tmpl_inputs.add_generation_prompt = add_generation_prompt;
-    tmpl_inputs.extra_context = extra_context;
+    tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
+    if (tools_override) {
+        tmpl_inputs.tools = *tools_override;
+    } else {
+        tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
+    }
+    tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
+    tmpl_inputs.extra_context = inputs.extra_context;
+    if (additional_context) {
+        tmpl_inputs.extra_context.merge_patch(*additional_context);
+    }
     // TODO: add flag to control date/time, if only for testing purposes.
     // tmpl_inputs.now = std::chrono::system_clock::now();
 
@@ -828,7 +838,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
         inputs.messages,
         "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
 
-    data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
     data.format = COMMON_CHAT_FORMAT_GENERIC;
     return data;
 }
@@ -904,7 +914,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.preserved_tokens = {
         "[TOOL_CALLS]",
     };
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     return data;
 }
@@ -934,7 +944,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
             adjusted_messages.push_back(msg);
         }
     }
-    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
     data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
     if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
         if (!inputs.enable_thinking) {
@@ -1122,7 +1132,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
     } else {
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
         {"date_string", format_time(inputs.now, "%d %b %Y")},
         {"tools_in_user_message", false},
         {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@@ -1187,7 +1197,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
 
 static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    auto prompt = apply(tmpl, inputs);
 
     // Hacks to fix the official (broken) prompt.
     // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@@ -1282,7 +1292,7 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
         {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
@@ -1338,7 +1348,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
     // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
     // If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -1465,7 +1475,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
 
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     // TODO: if (has_raw_python)
     return data;
 }
@@ -1498,14 +1508,15 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
 static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
-    json additional_context = {
+    json extra_context = json {
         {"enable_thinking", inputs.enable_thinking},
     };
+    extra_context.update(inputs.extra_context);
 
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
     data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
     if (string_ends_with(data.prompt, "<think>\n")) {
-        if (!inputs.enable_thinking) {
+        if (!extra_context["enable_thinking"]) {
             data.prompt += "</think>";
         } else {
             data.thinking_forced_open = true;
@@ -1691,7 +1702,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
 
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     data.grammar_lazy = false;
     if (!inputs.json_schema.is_null()) {
@@ -1722,6 +1733,12 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
+
+    params.extra_context = json::object();
+    for (auto el : inputs.chat_template_kwargs) {
+        params.extra_context[el.first] = json::parse(el.second);
+    }
+
     if (!inputs.json_schema.empty()) {
         params.json_schema = json::parse(inputs.json_schema);
     }

common/chat.h

@@ -7,6 +7,7 @@
 #include <chrono>
 #include <string>
 #include <vector>
+#include <map>
 
 struct common_chat_templates;
 
@@ -125,6 +126,7 @@ struct common_chat_templates_inputs {
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    std::map<std::string, std::string> chat_template_kwargs;
 };
 
 struct common_chat_params {

common/common.cpp

@@ -1005,15 +1005,21 @@ 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 +1163,7 @@ 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

@@ -8,6 +8,7 @@
 #include <string>
 #include <string_view>
 #include <vector>
+#include <map>
 #include <sstream>
 
 #ifdef _WIN32
@@ -80,6 +81,7 @@ enum llama_example {
     LLAMA_EXAMPLE_LOOKUP,
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
+    LLAMA_EXAMPLE_DIFFUSION,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -176,7 +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;
@@ -216,6 +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
@@ -267,6 +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;
 
@@ -329,6 +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
@@ -369,6 +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;
@@ -381,6 +395,8 @@ struct common_params {
     std::string ssl_file_key  = "";                                                                         // NOLINT
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
+    std::map<std::string, std::string> default_template_kwargs;
+
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
     bool endpoint_slots   = false;

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-s390x.md

@@ -16,7 +16,7 @@ cd llama.cpp
 
 ## CPU Build with BLAS
 
-Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements.
+Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. Make sure to have OpenBLAS installed in your environment.
 
 ```bash
 cmake -S . -B build             \
@@ -82,12 +82,18 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
 1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)**
 
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
     You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
 
-    These models and their respective tokenizers are verified to run correctly on IBM Z & LinuxONE.
+    These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system.
 
 2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
 
+    ![File Type - safetensors](https://img.shields.io/badge/File_Type-safetensors-da1e28)
+
+    The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
+
     ```bash
     python3 convert_hf_to_gguf.py \
         --outfile model-name-be.f16.gguf \
@@ -108,6 +114,10 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
 3. **Convert existing GGUF Little-Endian model to Big-Endian**
 
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
+    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
+
     ```bash
     python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
     ```
@@ -163,6 +173,22 @@ It is strongly recommended to disable SMT via the kernel boot parameters as it n
 
 IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS.
 
+## Frequently Asked Questions (FAQ)
+
+1. I'm getting the following error message while trying to load a model: `gguf_init_from_file_impl: failed to load model: this GGUF file version 50331648 is extremely large, is there a mismatch between the host and model endianness?`
+
+    Answer: Please ensure that the model you have downloaded/converted is GGUFv3 Big-Endian. These models are usually denoted with the `-be` suffix, i.e., `granite-3.3-2b-instruct-be.F16.gguf`.
+
+    You may refer to the [Getting GGUF Models](#getting-gguf-models) section to manually convert a `safetensors` model to `GGUF` Big Endian.
+
+2. I'm getting extremely poor performance when running inference on a model
+
+    Answer: Please refer to the [Appendix B: SIMD Support Matrix](#appendix-b-simd-support-matrix) to check if your model quantization is supported by SIMD acceleration.
+
+3. I'm building on IBM z17 and getting the following error messages: `invalid switch -march=z17`
+
+    Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
+
 ## Getting Help on IBM Z & LinuxONE
 
 1. **Bugs, Feature Requests**
@@ -172,3 +198,49 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 2. **Other Questions**
 
     Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com).
+
+## Appendix A: Hardware Support Matrix
+
+|         | Support | Minimum Compiler Version |
+| ------- | ------- | ------------------------ |
+| IBM z15 | ✅      |                          |
+| IBM z16 | ✅      |                          |
+| IBM z17 | ✅      | GCC 15.1.0               |
+
+-   ✅ - supported and verified to run as intended
+-   🚫 - unsupported, we are unlikely able to provide support
+
+## Appendix B: SIMD Support Matrix
+
+|            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
+| ---------- | ----------- | ---- | ---- | ----- |
+| FP32       | ✅          | ✅   | ❓   | ❓    |
+| FP16       | ✅          | ✅   | ❓   | ❓    |
+| BF16       | 🚫          | 🚫   | ❓   | ❓    |
+| Q4_0       | ✅          | ✅   | ❓   | ❓    |
+| Q4_1       | ✅          | ✅   | ❓   | ❓    |
+| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
+| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
+| Q8_0       | ✅          | ✅   | ❓   | ❓    |
+| Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
+| Q3_K       | ✅          | ✅   | ❓   | ❓    |
+| Q4_K       | ✅          | ✅   | ❓   | ❓    |
+| Q5_K       | ✅          | ✅   | ❓   | ❓    |
+| Q6_K       | ✅          | ✅   | ❓   | ❓    |
+| TQ1_0      | 🚫          | 🚫   | ❓   | ❓    |
+| TQ2_0      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XS     | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_M      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ4_NL     | ✅          | ✅   | ❓   | ❓    |
+| IQ4_XS     | ✅          | ✅   | ❓   | ❓    |
+| FP32->FP16 | 🚫          | ✅   | ❓   | ❓    |
+| FP16->FP32 | 🚫          | ✅   | ❓   | ❓    |
+
+-   ✅ - acceleration available
+-   🚫 - acceleration unavailable, will still run using scalar implementation
+-   ❓ - acceleration unknown, please contribute if you can test it yourself

docs/build.md

@@ -557,6 +557,23 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
+## WebGPU [In Progress]
+
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
+
+In the llama.cpp directory, build with CMake:
+
+```
+cmake -B build -DGGML_WEBGPU=ON
+cmake --build build --config Release
+```
+
+### Browser Support
+
+WebGPU allows cross-platform access to the GPU from supported browsers. We utilize [Emscripten](https://emscripten.org/) to compile ggml's WebGPU backend to WebAssembly. Emscripten does not officially support WebGPU bindings yet, but Dawn currently maintains its own WebGPU bindings called emdawnwebgpu.
+
+Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/src/emdawnwebgpu/) to download or build the emdawnwebgpu package (Note that it might be safer to build the emdawbwebgpu package locally, so that it stays in sync with the version of Dawn you have installed above). When building using CMake, the path to the emdawnwebgpu port file needs to be set with the flag `EMDAWNWEBGPU_DIR`.
+
 ## IBM Z & LinuxONE
 
 To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)

docs/development/HOWTO-add-model.md

@@ -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/docker.md

@@ -25,6 +25,9 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 

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/Miku.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 
 AI_NAME="${AI_NAME:-Miku}"

examples/chat-13B.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -e
 

examples/chat-persistent.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -euo pipefail
 

examples/chat-vicuna.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -e
 

examples/chat.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #
 # Temporary script - will be removed in the future

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

@@ -55,6 +55,8 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                         v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
                     } else if (type == GGML_TYPE_F32) {
                         v = *(float *) &data[i];
+                    } else if (type == GGML_TYPE_I64) {
+                        v = (float) *(int64_t *) &data[i];
                     } else if (type == GGML_TYPE_I32) {
                         v = (float) *(int32_t *) &data[i];
                     } else if (type == GGML_TYPE_I16) {
@@ -134,6 +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/jeopardy/jeopardy.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 
 MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin

examples/parallel/parallel.cpp

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

examples/reason-act.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 cd `dirname $0`
 cd ..

examples/server-llama2-13B.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -e
 

examples/simple-chat/simple-chat.cpp

@@ -113,15 +113,16 @@ int main(int argc, char ** argv) {
         while (true) {
             // check if we have enough space in the context to evaluate this batch
             int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0);
+            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) + 1;
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");
                 exit(0);
             }
 
-            if (llama_decode(ctx, batch)) {
-                GGML_ABORT("failed to decode\n");
+            int ret = llama_decode(ctx, batch);
+            if (ret != 0) {
+                GGML_ABORT("failed to decode, ret = %d\n", ret);
             }
 
             // sample the next token

examples/sycl/build.sh

@@ -1,4 +1,4 @@
-
+#!/usr/bin/env bash
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT

examples/sycl/run-llama2.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation

examples/sycl/run-llama3.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #  MIT license
 #  Copyright (C) 2025 Intel Corporation

examples/ts-type-to-grammar.sh

@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
 # python examples/json_schema_to_grammar.py https://json.schemastore.org/tsconfig.json

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-backend.h

@@ -339,7 +339,7 @@ extern "C" {
     typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node);
 
     // Tensor initialization
     GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

ggml/include/ggml-cpu.h

@@ -134,6 +134,7 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

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

@@ -314,6 +314,13 @@
 extern "C" {
 #endif
 
+    // Function type used in fatal error callbacks
+    typedef void (*ggml_abort_callback_t)(const char * error_message);
+
+    // Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
+    // Returns the old callback for chaining
+    GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
+
     GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
     GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
 
@@ -470,6 +477,7 @@ extern "C" {
         GGML_OP_TRANSPOSE,
         GGML_OP_GET_ROWS,
         GGML_OP_GET_ROWS_BACK,
+        GGML_OP_SET_ROWS,
         GGML_OP_DIAG,
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
@@ -481,12 +489,13 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         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,
@@ -519,6 +528,8 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
 
+        GGML_OP_GLU,
+
         GGML_OP_COUNT,
     };
 
@@ -542,6 +553,16 @@ extern "C" {
         GGML_UNARY_OP_COUNT,
     };
 
+    enum ggml_glu_op {
+        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,
+    };
+
     enum ggml_object_type {
         GGML_OBJECT_TYPE_TENSOR,
         GGML_OBJECT_TYPE_GRAPH,
@@ -627,6 +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);
@@ -657,6 +681,7 @@ extern "C" {
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_glu_op_name(enum ggml_glu_op op);
     GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
@@ -687,6 +712,9 @@ extern "C" {
     // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
     GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
 
+    // true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
+    GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
+
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -758,6 +786,7 @@ extern "C" {
     GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
     GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+    GGML_API enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor);
 
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
@@ -1086,6 +1115,89 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // gated linear unit ops
+    // A: n columns, r rows,
+    // result is n / 2 columns, r rows,
+    // expects gate in second half of row, unless swapped is true
+    GGML_API struct ggml_tensor * ggml_glu(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_glu_op     op,
+             bool                 swapped);
+
+    GGML_API struct ggml_tensor * ggml_reglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_reglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_swapped(
+            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(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             struct ggml_tensor * b,
+             enum ggml_glu_op     op);
+
+    GGML_API struct ggml_tensor * ggml_reglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_split(
+            struct ggml_context * ctx,
+            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,
@@ -1185,6 +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,
@@ -1375,6 +1500,23 @@ extern "C" {
             struct ggml_tensor  * b,  // row indices
             struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape
 
+    // a TD  [n_embd, ne1,    ne2,    ne3]
+    // b TS  [n_embd, n_rows, ne02,   ne03] | ne02 == ne2, ne03 == ne3
+    // c I64 [n_rows, ne11,   ne12,   1]    | c[i] in [0, ne1)
+    //
+    // undefined behavior if destination rows overlap
+    //
+    // broadcast:
+    //   ne2 % ne11 == 0
+    //   ne3 % ne12 == 0
+    //
+    // return view(a)
+    GGML_API struct ggml_tensor * ggml_set_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // destination
+            struct ggml_tensor  * b,  // source
+            struct ggml_tensor  * c); // row indices
+
     GGML_API struct ggml_tensor * ggml_diag(
         struct ggml_context     * ctx,
         struct ggml_tensor      * a);
@@ -1412,8 +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,
@@ -1723,6 +1871,17 @@ extern "C" {
             struct ggml_tensor  * b,
             int                   stride);
 
+    GGML_API struct ggml_tensor * ggml_conv_2d_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel [KW, KH, IC, OC]
+            struct ggml_tensor  * b,   // input data [W, H, C, N]
+            int                   s0,  // stride dimension 0
+            int                   s1,  // stride dimension 1
+            int                   p0,  // padding dimension 0
+            int                   p1,  // padding dimension 1
+            int                   d0,  // dilation dimension 0
+            int                   d1); // dilation dimension 1
+
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,
@@ -1765,6 +1924,12 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
+
+        GGML_SCALE_MODE_COUNT
+    };
+
+    enum ggml_scale_flag {
+        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
     };
 
     // interpolate
@@ -1777,14 +1942,26 @@ extern "C" {
 
     // interpolate
     // interpolate scale to specified dimensions
-    GGML_API struct ggml_tensor * ggml_upscale_ext(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
             int                   ne3,
-            enum ggml_scale_mode  mode);
+            enum ggml_scale_mode  mode),
+        "use ggml_interpolate instead");
+
+    // Up- or downsamples the input to the specified size.
+    // 2D scale modes (eg. bilinear) are applied to the first two dimensions.
+    GGML_API struct ggml_tensor * ggml_interpolate(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            uint32_t              mode); // ggml_scale_mode [ | ggml_scale_flag...]
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
@@ -1847,11 +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,
@@ -1890,7 +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");
 
@@ -817,8 +802,9 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         }
         if (sched->debug > 1) {
             ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
+                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
+                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
@@ -1826,7 +1812,7 @@ void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
     ggml_free(copy.ctx_unallocated);
 }
 
-bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
+bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node) {
     struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
     if (copy.buffer == NULL) {
         return false;
@@ -1837,28 +1823,45 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 
     assert(g1->n_nodes == g2->n_nodes);
 
-    for (int i = 0; i < g1->n_nodes; i++) {
-        struct ggml_tensor * t1 = g1->nodes[i];
-        struct ggml_tensor * t2 = g2->nodes[i];
+    if (test_node != nullptr) {
+        // Compute the whole graph and only test the output for a specific tensor
+        ggml_backend_graph_compute(backend1, g1);
+        ggml_backend_graph_compute(backend2, g2);
 
-        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
-
-        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
-        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
-
-        ggml_backend_graph_compute(backend1, &g1v);
-        ggml_backend_graph_compute(backend2, &g2v);
-
-        if (ggml_is_view_op(t1->op)) {
-            continue;
+        int test_node_idx = -1;
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            if (t1 == test_node) {
+                test_node_idx = i;
+                break;
+            }
         }
+        GGML_ASSERT(test_node_idx != -1);
 
-        // compare results, calculate rms etc
-        if (!callback(i, t1, t2, user_data)) {
-            break;
+        callback(test_node_idx, g1->nodes[test_node_idx], g2->nodes[test_node_idx], user_data);
+    } else {
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            struct ggml_tensor * t2 = g2->nodes[i];
+
+            assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+            struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+            struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+            ggml_backend_graph_compute(backend1, &g1v);
+            ggml_backend_graph_compute(backend2, &g2v);
+
+            if (ggml_is_view_op(t1->op)) {
+                continue;
+            }
+
+            // compare results, calculate rms etc
+            if (!callback(i, t1, t2, user_data)) {
+                break;
+            }
         }
     }
-
     ggml_backend_graph_copy_free(copy);
 
     return true;

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -65,8 +65,9 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
-#include <aclnnop/aclnn_grouped_matmul_v2.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);
 }
 
 /**
@@ -2654,6 +2656,67 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
     }
 
+#ifdef ASCEND_310P
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    if (src0->type == GGML_TYPE_F16) {
+        src0_row.type = GGML_TYPE_F32;
+    }
+
+    // src0_row [D, M, 1, 1] weight without permute
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[0] = ori_src0_nb[0];
+    src0_row.nb[1] = ori_src0_nb[1];
+    src0_row.nb[2] = ori_src0_nb[1];
+    src0_row.nb[3] = ori_src0_nb[1];
+
+    // src1_row [D, 1, 1, 1] -> input
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    // dst_row [M, 1, 1, 1] -> out
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+
+    //create weight for one row
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            src0_row.data = src0_tmp_ptr;
+            src1_row.data = src1_tmp_ptr;
+            dst_row.data = dst_tmp_ptr;
+            dst_row.src[0] = &src0_row;
+            dst_row.src[1] = &src1_row;
+
+            ggml_cann_mul_mat(ctx, &dst_row);
+        }
+    }
+    return;
+#endif
+
     std::vector<aclTensor*> src0_tensor_vec;
     std::vector<aclTensor*> src1_tensor_vec;
     std::vector<aclTensor*> dst_tensor_vec;
@@ -2701,9 +2764,9 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
     }
 
     size_t GROUP_SIZE = 128;
-    // GroupedMatmulV2 required tensor_list.size < 128
+    // GroupedMatmulV3 required tensor_list.size < 128
     for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
-        // split and call GroupedMatmulV2
+        // split and call GroupedMatmulV3
         size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
         std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
         std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@@ -2713,7 +2776,7 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
         aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
 
-        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
             nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
 
         ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);

ggml/src/ggml-cann/common.h

@@ -359,7 +359,7 @@ struct ggml_backend_cann_context {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
 
-        bool async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
+        async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
     }

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/CMakeLists.txt

@@ -5,7 +5,7 @@ function(ggml_add_cpu_backend_features cpu_name arch)
     # build, using set_source_files_properties() to set the arch flags is not possible
     set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
     add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
-    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include)
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
     target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
     set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
@@ -589,4 +589,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     if (EMSCRIPTEN)
         set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
     endif()
+
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
+    endif()
 endfunction()

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

@@ -195,6 +195,7 @@ typedef pthread_t ggml_thread_t;
 
 static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32] = {
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp32,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
         .vec_dot_type             = GGML_TYPE_F32,
         .nrows                    = 1,
@@ -1192,7 +1193,7 @@ static void ggml_compute_forward_mul_mat_one_chunk(
     }
 }
 
-static void ggml_compute_forward_mul_mat(
+void ggml_compute_forward_mul_mat(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
 
@@ -1817,6 +1818,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_get_rows_back(params, tensor);
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                ggml_compute_forward_set_rows(params, tensor);
+            } break;
         case GGML_OP_DIAG:
             {
                 ggml_compute_forward_diag(params, tensor);
@@ -1861,6 +1866,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_im2col_back_f32(params, tensor);
             } break;
+        case GGML_OP_CONV_2D:
+            {
+                ggml_compute_forward_conv_2d(params, tensor);
+            } break;
         case GGML_OP_CONV_2D_DW:
             {
                 ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -1944,6 +1953,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_unary(params, tensor);
             } break;
+        case GGML_OP_GLU:
+            {
+                ggml_compute_forward_glu(params, tensor);
+            } break;
         case GGML_OP_GET_REL_POS:
             {
                 ggml_compute_forward_get_rel_pos(params, tensor);
@@ -2154,6 +2167,20 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     GGML_ABORT("fatal error");
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(node)) {
+                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;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+            break;
         case GGML_OP_SILU_BACK:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -2170,6 +2197,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_GET_ROWS:
+        case GGML_OP_SET_ROWS:
             {
                 // FIXME: get_rows can use additional threads, but the cost of launching additional threads
                 // decreases performance with GPU offloading
@@ -2206,6 +2234,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2724,6 +2753,10 @@ struct ggml_cplan ggml_graph_plan(
                             GGML_ABORT("fatal error");
                         }
                     } break;
+                case GGML_OP_CONV_2D:
+                    {
+                        cur = GGML_IM2COL_WORK_SIZE;
+                    } break;
                 case GGML_OP_CONV_TRANSPOSE_2D:
                     {
                         const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -3124,6 +3157,10 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+void ggml_cpu_fp32_to_fp32(const float * x, float * y, int64_t n) {
+    memcpy(y, x, n * sizeof(float));
+}
+
 void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
     int64_t i = 0;
 #if defined(__F16C__)

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

@@ -416,6 +416,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
 
     switch (op->op) {
         case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
             return
                 op->type != GGML_TYPE_IQ3_XXS &&
                 op->type != GGML_TYPE_IQ3_S   &&

ggml/src/ggml-cpu/ops.h

@@ -20,6 +20,9 @@
 
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
+// Work buffer size for im2col operations in CONV2D
+#define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -53,6 +56,7 @@ void ggml_compute_forward_permute(const struct ggml_compute_params * params, str
 void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -64,6 +68,7 @@ void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struc
 void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -93,6 +98,7 @@ void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, st
 void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_glu(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -105,6 +111,7 @@ void ggml_compute_forward_custom(const struct ggml_compute_params * params, stru
 void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus
 }

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]));
@@ -254,6 +257,30 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
     }
 }
 
+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g) {
+    int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_silu(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_silu(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        _mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = ggml_silu_f32(x[i]) * g[i];
+    }
+}
+
 ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
     int i = 0;
     ggml_float sum = 0;

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)
@@ -905,6 +944,100 @@ inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, con
     }
 }
 
+inline static void ggml_vec_reglu_f32 (const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = (x[i] > 0.f) ? x[i] * g[i] : 0.f;
+    }
+}
+
+inline static void ggml_vec_reglu_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 v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v * GGML_CPU_FP16_TO_FP32(g[i]) : 0.f);
+    }
+}
+
+#ifdef GGML_GELU_FP16
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+    uint16_t t;
+    for (int i = 0; i < n; ++i) {
+        if (x[i] <= -10.0f) {
+            y[i] = 0.0f;
+        } else if (x[i] >= 10.0f) {
+            y[i] = x[i] * g[i];
+        } else {
+            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_f16[t]) * g[i];
+        }
+    }
+}
+#else
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_gelu_f32(x[i]) * g[i];
+    }
+}
+#endif
+
+inline static void ggml_vec_geglu_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_f16[i16[i]]) * v);
+    }
+}
+
+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g);
+
+inline static void ggml_vec_swiglu_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 v = GGML_CPU_FP16_TO_FP32(x[i]);
+        float w = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v/(1.0f + expf(-v))) * w);
+    }
+}
+
+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/convert.cu

@@ -728,3 +728,25 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
             return nullptr;
     }
 }
+
+to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float, nv_bfloat16>;
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half, nv_bfloat16>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half, float>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cuda<nv_bfloat16, float>;
+        default:
+            return nullptr;
+    }
+}

ggml/src/ggml-cuda/convert.cuh

@@ -22,5 +22,10 @@ using to_t_nc_cuda_t = void (*)(const void * x, T * y,
     int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
     int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
 
+typedef to_t_nc_cuda_t<float> to_fp32_nc_cuda_t;
 typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
+typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
+
+to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
 to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
+to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);

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>
@@ -1749,7 +1750,7 @@ static void ggml_cuda_op_mul_mat(
 }
 
 static __global__ void k_compute_batched_ptrs(
-        const half * src0_as_f16, const half * src1_as_f16, char * dst,
+        const void * src0_as_f16, const void * src1_as_f16, char * dst,
         const void ** ptrs_src, void ** ptrs_dst,
         int64_t ne12, int64_t ne13,
         int64_t ne23,
@@ -1772,83 +1773,131 @@ static __global__ void k_compute_batched_ptrs(
     ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
 }
 
-static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+// Type traits for mapping ggml types to CUDA/cuBLAS types
+template<ggml_type T>
+struct batched_mul_mat_traits;
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_F32> {
+    using cuda_type = float;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
+    static inline const cudaDataType_t data_type = CUDA_R_32F;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_F32;
+    static inline const float alpha = 1.0f;
+    static inline const float beta = 0.0f;
+    static inline const void* get_alpha() { static const float val = alpha; return &val; }
+    static inline const void* get_beta() { static const float val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp32_nc_cuda(src_type); }
+};
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_BF16> {
+    using cuda_type = nv_bfloat16;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
+    static inline const cudaDataType_t data_type = CUDA_R_16BF;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_BF16;
+    static inline const float alpha = 1.0f;
+    static inline const float beta = 0.0f;
+    static inline const void* get_alpha() { static const float val = alpha; return &val; }
+    static inline const void* get_beta() { static const float val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_bf16_nc_cuda(src_type); }
+};
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_F16> {
+    using cuda_type = half;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_16F;
+    static inline const cudaDataType_t data_type = CUDA_R_16F;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_F16;
+    static inline const half alpha = 1.0;
+    static inline const half beta = 0.0;
+    static inline const void* get_alpha() { static const half val = alpha; return &val; }
+    static inline const void* get_beta() { static const half val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp16_nc_cuda(src_type); }
+};
+
+template<ggml_type src0_type>
+static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    using traits = batched_mul_mat_traits<src0_type>;
+    using cuda_t = typename traits::cuda_type;
+
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
-
     GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == src0_type);
+    GGML_ASSERT(ggml_is_contiguous(dst));
 
     // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
     // As long as dst is contiguous this does not matter though.
-    GGML_ASSERT(ggml_is_contiguous(dst));
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t ne_dst = ggml_nelements(dst);
-
     cudaStream_t main_stream = ctx.stream();
-
     CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
 
-    const half * src0_f16 = (const half *) src0->data;
     float * dst_ddf = (float *) dst->data;
-
-    const half * src1_f16 = (const half *) src1->data;
     const size_t ts_src1 = ggml_type_size(src1->type);
     GGML_ASSERT(nb10 == ts_src1);
     int64_t s11 = nb11 / ts_src1;
     int64_t s12 = nb12 / ts_src1;
     int64_t s13 = nb13 / ts_src1;
-    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
 
-    // convert src1 to fp16
-    if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
+    const cuda_t * src0_ptr = nullptr;
+    const cuda_t * src1_ptr = nullptr;
+
+    ggml_cuda_pool_alloc<cuda_t> src0_alloc(ctx.pool());
+    ggml_cuda_pool_alloc<cuda_t> src1_alloc(ctx.pool());
+
+    // Handle src0
+    src0_ptr = (const cuda_t *) src0->data;
+
+    // Handle src1 - convert if necessary
+    if (src1->type == src0_type) {
+        src1_ptr = (const cuda_t *) src1->data;
+    } else {
+        // Convert src1 to target type using traits conversion functions
         const int64_t ne_src1 = ggml_nelements(src1);
-        src1_f16_alloc.alloc(ne_src1);
-        GGML_ASSERT(to_fp16_cuda != nullptr);
+        src1_alloc.alloc(ne_src1);
 
-        to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
-
-        src1_f16 = src1_f16_alloc.get();
+        const auto convert_func = traits::get_nc_converter(src1->type);
+        GGML_ASSERT(convert_func != nullptr);
+        convert_func(src1->data, src1_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+        src1_ptr = src1_alloc.get();
         s11 = ne10;
         s12 = ne11*s11;
         s13 = ne12*s12;
     }
 
-    ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
+    // Setup destination buffer
+    ggml_cuda_pool_alloc<cuda_t> dst_temp(ctx.pool());
     char * dst_t;
-
-    cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
-    cudaDataType_t      cu_data_type    = CUDA_R_16F;
-
-    // dst strides
     size_t nbd2 = dst->nb[2];
     size_t nbd3 = dst->nb[3];
 
-    const half  alpha_f16 = 1.0f;
-    const half  beta_f16  = 0.0f;
-
+    cublasComputeType_t cu_compute_type = traits::compute_type;
+    cudaDataType_t cu_data_type = traits::data_type;
+    cudaDataType_t cu_data_type_a = traits::data_type;
+    cudaDataType_t cu_data_type_b = traits::data_type;
+    const void * alpha = traits::get_alpha();
+    const void * beta = traits::get_beta();
     const float alpha_f32 = 1.0f;
-    const float beta_f32  = 0.0f;
-
-    const void * alpha = &alpha_f16;
-    const void * beta  = &beta_f16;
+    const float beta_f32 = 0.0f;
 
     if (dst->op_params[0] == GGML_PREC_DEFAULT) {
-        dst_t = (char *) dst_f16.alloc(ne_dst);
-
-        nbd2 /= sizeof(float) / sizeof(half);
-        nbd3 /= sizeof(float) / sizeof(half);
+        if constexpr (src0_type == GGML_TYPE_F32) {
+            dst_t = (char *) dst_ddf;  // Direct F32 output
+        } else {
+            dst_t = (char *) dst_temp.alloc(ne_dst);
+            nbd2 /= sizeof(float) / sizeof(cuda_t);
+            nbd3 /= sizeof(float) / sizeof(cuda_t);
+        }
     } else {
         dst_t = (char *) dst_ddf;
-
         cu_compute_type = CUBLAS_COMPUTE_32F;
-        cu_data_type    = CUDA_R_32F;
-
+        cu_data_type = CUDA_R_32F;
         alpha = &alpha_f32;
-        beta  = &beta_f32;
+        beta = &beta_f32;
     }
 
     int id = ggml_cuda_get_device();
@@ -1856,7 +1905,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
         cu_compute_type = CUBLAS_COMPUTE_32F;
         alpha = &alpha_f32;
-        beta  = &beta_f32;
+        beta = &beta_f32;
     }
 
     GGML_ASSERT(ne12 % ne02 == 0);
@@ -1866,35 +1915,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     const int64_t r2 = ne12/ne02;
     const int64_t r3 = ne13/ne03;
 
-#if 0
-    // use cublasGemmEx
-    {
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                int i03 = i13 / r3;
-                int i02 = i12 / r2;
-
-                CUBLAS_CHECK(
-                cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
-                    ne01, ne11, ne10,
-                    alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F,   nb01/sizeof(half),
-                                          src1_f16 + i13*s13  + i12*s12,  CUDA_R_16F,   s11,
-                    beta,  (      char *)    dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
-                    cu_compute_type,
-                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-            }
-        }
-    }
-#else
     if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00, // strideA
-                       src1_f16, CUDA_R_16F,   s11,       s12,       // strideB
-                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
+                alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
+                       src1_ptr, cu_data_type_b, s11,       s12,       // strideB
+                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0,   // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1905,34 +1934,55 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
         ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
 
+        size_t src1_stride_size = sizeof(cuda_t);
+
         dim3 block_dims(ne13, ne12);
         k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
-                src0_f16, src1_f16, dst_t,
+                src0_ptr, src1_ptr, dst_t,
                 ptrs_src.get(), ptrs_dst.get(),
                 ne12, ne13,
                 ne23,
                 nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
-                src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
+                (src1->type == src0_type) ? nb12 : s12*src1_stride_size,
+                (src1->type == src0_type) ? nb13 : s13*src1_stride_size,
                 nbd2, nbd3,
                 r2, r3);
+
         CUDA_CHECK(cudaGetLastError());
 
         CUBLAS_CHECK(
         cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   s11,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
+                alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, nb01/nb00,
+                       (const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type,   ne0,
                 ne23,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     }
-#endif
 
-    if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type == CUDA_R_16F) {
-        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-        to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
+    // Convert output back to F32 if needed
+    if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type != CUDA_R_32F) {
+        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(traits::ggml_type_val);
+        to_fp32_cuda(dst_temp.get(), dst_ddf, ne_dst, main_stream);
+    }
+}
+
+static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F32);
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F32>(ctx, src0, src1, dst);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_BF16>(ctx, src0, src1, dst);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F16>(ctx, src0, src1, dst);
+            break;
+        default:
+            GGML_ABORT("Unsupported type");
     }
 }
 
@@ -1984,6 +2034,12 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
+    //TODO update for generic tensor parallelism
+    const int cc                     = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    bool use_batched_cublas_f16  = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
+    bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
+    bool use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;
+
     if (!split && use_mul_mat_vec) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
@@ -1992,8 +2048,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
     } else if (!split && use_mul_mat_q) {
         ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
-            !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && (use_batched_cublas_f16 || use_batched_cublas_bf16 || use_batched_cublas_f32)
+        && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_mul_mat_vec) {
@@ -2175,6 +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;
@@ -2244,6 +2303,30 @@ 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;
+            }
+            break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(dst)) {
+                case GGML_GLU_OP_REGLU:
+                    ggml_cuda_op_reglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                    ggml_cuda_op_geglu(ctx, dst);
+                    break;
+                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;
             }
@@ -2507,6 +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];
 
@@ -2528,9 +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]);
@@ -3036,11 +3125,24 @@ 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;
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous_1(op->src[0]);
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -3112,6 +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:
@@ -3126,6 +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;
@@ -3241,12 +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: {
@@ -3271,7 +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:
@@ -3295,9 +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;
             }
@@ -3310,6 +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 +200,106 @@ 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>
+static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1) {
+    const int64_t i = int64_t(blockDim.x)*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    // perform base op and multiply with gate (either offset in same tensor or a separate one)
+    const int64_t j0 = (i / n) * o0 + (i % n);
+    const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+
+    dst[i] = (T)(op((float)x[j0]) * (float)g[j1]);
+}
+
+template <float (*op)(float), typename T>
+static void unary_gated_cuda(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, cudaStream_t stream) {
+    const int64_t num_blocks = (k + CUDA_GLU_BLOCK_SIZE - 1) / CUDA_GLU_BLOCK_SIZE;
+    unary_gated_op_kernel<op><<<num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream>>>(x, g, dst, k, n, o0, o1);
+}
+
+template <float (*op)(float)>
+void ggml_cuda_op_unary_gated(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    void * src0_d = src0->data;
+    void * src1_d = src1 ? src1->data : src0->data;
+    const int64_t src0_o = src0->nb[1];
+    const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+    void * dst_d = dst->data;
+    const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(src0->nb[0] == ggml_element_size(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == ggml_nrows(src0));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
+        GGML_ASSERT(src1->ne[0] == nc);
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int32_t swapped = ((const int32_t *) dst->op_params)[1];
+
+    if (src0->type == GGML_TYPE_F16) {
+        half * src0_p = (half *) src0_d;
+        half * src1_p = (half *) src1_d;
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        unary_gated_cuda<op>(src0_p, src1_p, (half *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(half), src1_o / sizeof(half), stream);
+    } else {
+        float * src0_p = (float *) src0_d;
+        float * src1_p = (float *) src1_d;
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        unary_gated_cuda<op>(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), stream);
+    }
+}
+
+void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_relu>(ctx, dst);
+}
+
+void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu>(ctx, dst);
+}
+
+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

@@ -15,6 +15,7 @@
 #define CUDA_SQRT_BLOCK_SIZE 256
 #define CUDA_SIN_BLOCK_SIZE 256
 #define CUDA_COS_BLOCK_SIZE 256
+#define CUDA_GLU_BLOCK_SIZE 256
 
 void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
@@ -57,3 +58,15 @@ void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 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__)