metal : port new kernels into the split sources

Move the kernels added on master after the split (lightning indexer, DSv4 hyper-connections, silu_back, f16 bin ops, TQ2_0, the flash-attn KV dequantization pass, rope offset/inplace, ssm_scan rollback, packed q8_0 dequantization and the tensor-API mat-mat K clamp) into the corresponding kernels/*.metal sources. Copied verbatim, no functional change.
ggml-metal: FWHT kernel for metal backend (#25924 )
2026-08-24 05:32:32 +02:00 · 2026-08-22 13:40:28 +08:00 · 2026-08-22 13:40:28 +08:00 · 2026-08-22 13:40:28 +08:00 · 2026-08-22 13:40:28 +08:00 · 2026-08-22 13:40:28 +08:00
440 changed files with 33239 additions and 28148 deletions
@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2.1
-ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04

 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.36.3
-ARG IGC_VERSION_FULL=2_2.36.3+21719
-ARG COMPUTE_RUNTIME_VERSION=26.22.38646.4
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.22.38646.4-0
+ARG IGC_VERSION=v2.38.2
+ARG IGC_VERSION_FULL=2_2.38.2+22051
+ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
 ARG IGDGMM_VERSION=22.10.0

 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.33.0
-ARG NPU_DRIVER_FULL=v1.33.0.20260529-26625960453
-ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
+ARG NPU_DRIVER_VERSION=v1.35.0
+ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
+ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1

 # Optional proxy build arguments
 ARG http_proxy=
@@ -170,7 +170,7 @@ RUN --mount=type=cache,target=/var/cache/intel-npu,sharing=locked \
    fi; \
    DEB=/var/cache/intel-npu/libze1_${LIBZE1_VERSION}_amd64.deb; \
    if [ ! -f "$DEB" ]; then \
-        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
    fi; \
    mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
    apt-get update; \
@@ -1,3 +1,4 @@
+# note: place this as the last step of the job, so the new cache is saved by "Post ccache" right after the old one is cleared
 name: "ccache-clear"
 description: "Delete all GitHub Actions caches matching a key prefix"
 inputs:
@@ -1,20 +0,0 @@
-name: "Linux - Setup Vulkan SDK"
-description: "Setup Vulkan SDK for Linux"
-inputs:
-  path:
-    description: "Installation path"
-    required: true
-  version:
-    description: "Vulkan SDK version"
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Setup Vulkan SDK
-      id: setup
-      uses: ./.github/actions/unarchive-tar
-      with:
-        url: https://sdk.lunarg.com/sdk/download/${{ inputs.version }}/linux/vulkan_sdk.tar.xz
-        path: ${{ inputs.path }}
-        strip: 1
@@ -6,8 +6,7 @@ inputs:
    required: true
  cuda_arch:
    description: "CUDA target architecture"
-    required: false
-    default: "x64"
+    required: true

 runs:
  using: "composite"
@@ -10,33 +10,6 @@ concurrency:
  cancel-in-progress: true

 jobs:
-  ubuntu-24-vulkan-cache:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Get latest Vulkan SDK version
-        id: vulkan_sdk_version
-        run: |
-          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
-
-      - name: Setup Cache
-        uses: actions/cache@v5
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
-      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
-        with:
-          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
-
  #ubuntu-24-spacemit-cache:
  #  runs-on: ubuntu-24.04

@@ -67,9 +40,9 @@ jobs:
    runs-on: ubuntu-24.04

    env:
-      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -96,8 +69,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -27,30 +27,26 @@ jobs:
          cmake --install build --prefix "$PREFIX" --config Release

          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          tclsh <<'EOF'
-          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
-          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          build_commit=$(git rev-parse --short HEAD | xargs)
+          build_number=$(git rev-list --count HEAD | xargs)

-          set cmakelists [read [open "CMakeLists.txt" r]]
-          regexp {set\(LLAMA_VERSION_MAJOR\s+(\d+)\)} $cmakelists -> major
-          regexp {set\(LLAMA_VERSION_MINOR\s+(\d+)\)} $cmakelists -> minor
-          regexp {set\(LLAMA_VERSION_PATCH\s+(\d+)\)} $cmakelists -> patch
-          set build(version) "$major.$minor.$patch"
+          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          build_version="$major.$minor.$patch"

-          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
-          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
-                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
-                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
+                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
+                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")

-          puts -nonewline "Checking llama-config.cmake version... "
-          foreach check $checks {
-              if {![regexp -expanded -- $check $llamaconfig]} {
-                  puts "\"$check\" failed!"
+          for check in "${checks[@]}"; do
+              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
+                  echo "Checking llama-config.cmake version... \"$check\" failed!"
                  exit 1
-              }
-          }
-          puts "success."
-          EOF
+              fi
+          done
+
+          echo "Checking llama-config.cmake version... success."

          cd examples/simple-cmake-pkg
          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
@@ -21,6 +21,7 @@ on:
    paths: [
      '.github/workflows/build-cpu.yml',
      '.github/workflows/build-cmake-pkg.yml',
+      'ggml/src/ggml-rpc/**',
      '**/CMakeLists.txt',
      '**/.cmake',
      '**/*.h',
@@ -96,8 +97,7 @@ jobs:
          cmake -B build \
            -DGGML_NATIVE=OFF \
            -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON \
-            -DGGML_NATIVE=OFF
+            -DGGML_RPC=ON
          time cmake --build build --config Release -j $(nproc)

      - name: Test
@@ -118,28 +118,25 @@ jobs:
          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

  windows:
+    name: windows / ${{ matrix.build }}
    runs-on: windows-2025

    env:
      OPENBLAS_VERSION: 0.3.23
      SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.357.0

    strategy:
      matrix:
        include:
          - build: 'x64-cpu-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'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
          - build: 'x64-openblas'
            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: 'x64-vulkan'
-            arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -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: 'arm64'
            arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON'

    steps:
      - name: Clone
@@ -167,15 +164,6 @@ jobs:
          $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
          & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll

-      - name: Install Vulkan SDK
-        id: get_vulkan
-        if: ${{ matrix.build == 'x64-vulkan' }}
-        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
-          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
-          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
-
      - name: Install Ninja
        id: install_ninja
        run: |
@@ -22,6 +22,7 @@ env:

 jobs:
  cuda:
+    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
    runs-on: windows-2022

    permissions:
@@ -29,7 +30,16 @@ jobs:

    strategy:
      matrix:
-        cuda: ['12.4', '13.3']
+        include:
+          - cuda: '12.4'
+            arch: x64
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
+          - cuda: '13.3'
+            arch: x64
+            defines: ''
+          - cuda: '13.4'
+            arch: arm64
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -39,12 +49,13 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

      - name: Install Cuda Toolkit
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -54,26 +65,21 @@ jobs:
      - name: Build
        id: cmake_build
        shell: cmd
-        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
        run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          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" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_NATIVE=OFF ^
            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
+            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

  hip:
    runs-on: windows-2022
@@ -39,8 +39,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -81,7 +81,7 @@ jobs:
        # TODO: fix and re-enable the `test-llama-archs` test below
        run: |
          cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000

      - name: Test (GPU)
        id: cmake_test_gpu
@@ -89,15 +89,15 @@ jobs:
        run: |
          cd ${{ github.workspace }}
          export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000

  openvino-windows-2022:
    runs-on: windows-2022

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -166,4 +166,4 @@ jobs:
          call "%OPENVINO_ROOT%\setupvars.bat"

          cd build
-          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000
@@ -1,66 +0,0 @@
-name: CI (rpc)
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-    paths: [
-      '.github/workflows/build-rpc.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
-
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-rpc.yml',
-      'ggml/src/ggml-rpc/**'
-    ]
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_ARG_LOG_COLORS: 1
-  LLAMA_ARG_LOG_PREFIX: 1
-  LLAMA_ARG_LOG_TIMESTAMPS: 1
-
-jobs:
-  ubuntu-24-rpc:
-    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libssl-dev ninja-build
-
-      - name: Build
-        id: cmake_build
-        run: |
-          cmake -B build \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_RPC=ON
-          time cmake --build build --config Release -j $(nproc)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose
@@ -288,8 +288,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -93,19 +93,13 @@ jobs:
        run: |
          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"

-      - name: Use Vulkan SDK Cache
-        uses: actions/cache@v5
-        id: cache-sdk
-        with:
-          path: ./vulkan_sdk
-          key: cache-gha-vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
-
      - name: Setup Vulkan SDK
-        if: steps.cache-sdk.outputs.cache-hit != 'true'
-        uses: ./.github/actions/linux-setup-vulkan
+        id: setup
+        uses: ./.github/actions/unarchive-tar
        with:
+          url: https://sdk.lunarg.com/sdk/download/${{ env.VULKAN_SDK_VERSION }}/linux/vulkan_sdk.tar.xz
          path: ./vulkan_sdk
-          version: ${{ env.VULKAN_SDK_VERSION }}
+          strip: 1

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -133,3 +127,56 @@ jobs:
          # This is using llvmpipe and runs slower than other backends
          # test-backend-ops is too slow on llvmpipe, skip it
          ctest -L main -E test-backend-ops --verbose --timeout 900
+
+  windows:
+    runs-on: windows-2025
+
+    env:
+      VULKAN_VERSION: 1.4.357.0
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: cpu-windows-2025-x64-vulkan
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        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
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -S . -B build -G "Ninja Multi-Config" `
+            -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake `
+            -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 `
+            -DLLAMA_BUILD_BORINGSSL=ON
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main -C Release --verbose --timeout 900
@@ -394,6 +394,11 @@ jobs:
    name: Create shared tags from digests
    needs: [prepare_matrices, push_to_registry, create_tag]
    runs-on: ubuntu-24.04
+    permissions:
+      contents: read
+      packages: write
+      id-token: write
+      attestations: write
    strategy:
      fail-fast: false
      matrix:
@@ -428,6 +433,7 @@ jobs:
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Create tags from digests
+        id: create_tags
        shell: bash
        run: |
          set -euo pipefail
@@ -439,6 +445,7 @@ jobs:
          SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
          BUILD_DATE="${{ steps.build_date.outputs.date }}"
          COMMIT_SHA="${{ steps.checkout.outputs.commit }}"
+          echo "image_repo=${IMAGE_REPO}" >> "$GITHUB_OUTPUT"
          TAGS="${{ matrix.config.tag }}"
          ARCHES="${{ matrix.config.arches }}"
          DIGEST_GLOB="/tmp/digests/*.tsv"
@@ -505,6 +512,16 @@ jobs:

              echo "Creating ${merged_versioned_tag} from ${refs[*]}"
              docker buildx imagetools create "${annotations[@]}" --tag "${merged_versioned_tag}" "${refs[@]}"
+
+              if [[ "$tag_name" == "${TAGS%% *}" ]]; then
+                  local digest
+                  digest="$(docker buildx imagetools inspect "${merged_versioned_tag}" --format '{{.Manifest.Digest}}')"
+                  if [[ ! "$digest" =~ ^sha256:[0-9a-f]{64}$ ]]; then
+                      echo "Invalid digest for ${merged_versioned_tag}: ${digest}" >&2
+                      exit 1
+                  fi
+                  echo "${image_type}_digest=${digest}" >> "$GITHUB_OUTPUT"
+              fi
          }

          for tag in $TAGS; do
@@ -528,3 +545,24 @@ jobs:
          done
        env:
          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
+      - name: Attest full image
+        if: ${{ matrix.config.full }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.full_digest }}
+
+      - name: Attest light image
+        if: ${{ matrix.config.light }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.light_digest }}
+
+      - name: Attest server image
+        if: ${{ matrix.config.server }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.server_digest }}
@@ -49,12 +49,77 @@ jobs:
          git push origin "${VERSION}"
          echo "Created and pushed tag ${VERSION}"

+      - name: Generate release description
+        id: desc
+        run: bash scripts/make-release-desc.sh "${{ steps.checks.outputs.version }}"
+        env:
+          GITHUB_REPOSITORY: ${{ github.repository }}
+
+      - name: Create nightly-tag.txt
+        id: nightly_tag_file
+        run: |
+          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
+          if [[ -z "${NIGHTLY_TAG}" ]]; then
+            echo "Warning: no nightly tag found for the release commit - nightly-tag.txt will not be created"
+            echo "create=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          echo "${NIGHTLY_TAG}" > nightly-tag.txt
+          echo "create=true" >> "$GITHUB_OUTPUT"
+          echo "nightly-tag.txt:"
+          cat nightly-tag.txt
+
+      - name: Create release
+        id: create_release
+        if: ${{ github.event.inputs.dry_run == 'false' }}
+        uses: ggml-org/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        with:
+          tag_name: ${{ steps.checks.outputs.version }}
+          prerelease: false
+          # TODO: enrich the body of the release with more information
+          body: |
+            ## Overview
+
+            New version has been released.
+
+            ${{ steps.desc.outputs.nightly }}
+
+            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
+
+            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
+
+            ## ${{ steps.desc.outputs.changelog_title }}
+
+            ${{ steps.desc.outputs.changelog }}
+
+      - name: Upload nightly-tag.txt
+        if: ${{ github.event.inputs.dry_run == 'false' && steps.nightly_tag_file.outputs.create == 'true' }}
+        uses: actions/github-script@v8
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            console.log('uploadReleaseAsset', 'nightly-tag.txt');
+            await github.rest.repos.uploadReleaseAsset({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              release_id: release_id,
+              name: 'nightly-tag.txt',
+              data: await fs.readFileSync('./nightly-tag.txt')
+            });
+
      - name: Dry run summary
        if: ${{ github.event.inputs.dry_run == 'true' }}
        run: |
          if [[ "${{ steps.checks.outputs.checks_passed }}" == "true" ]]; then
            echo "Dry run complete - all checks passed."
            echo "Would have created tag: ${{ steps.checks.outputs.version }}"
+            if [[ -n "${{ steps.desc.outputs.nightly_tag }}" ]]; then
+              echo "Would have uploaded nightly-tag.txt: ${{ steps.desc.outputs.nightly_tag }}"
+            fi
          else
            echo "::error::Dry run found release check failures. A release tag would not be created."
            exit 1
@@ -145,11 +145,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-${{ matrix.arch }}
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -166,6 +161,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-macos-${{ matrix.build }}.tar.gz
          name: llama-bin-macos-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-${{ matrix.arch }}
+
  ubuntu-cpu:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -231,12 +231,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        if: ${{ matrix.build != 's390x' }}
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-cpu
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -253,6 +247,12 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
          name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        if: ${{ matrix.build != 's390x' }}
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-cpu
+
  ubuntu-vulkan:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -318,11 +318,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-vulkan
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -339,6 +334,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
          name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-vulkan
+
  android-arm64:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -446,8 +446,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -512,11 +512,6 @@ jobs:
            ${{ env.CMAKE_ARGS }}
          cmake --build build/ReleaseOV --config Release --parallel

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-openvino-release-no-preset-v1
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -551,6 +546,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
          name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-openvino-release-no-preset-v1
+
  windows-openvino:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -562,8 +562,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -637,11 +637,6 @@ jobs:

          cmake --build build\ReleaseOV --config Release -- /m

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-openvino
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -680,7 +675,13 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
          name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-openvino
+
  windows-cpu:
+    name: windows-cpu / ${{ matrix.arch }}
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

@@ -728,18 +729,13 @@ jobs:
            -DGGML_BACKEND_DL=ON ^
            -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
            -DGGML_OPENMP=ON ^
+            -DGGML_OPENMP_FETCH=ON ^
            ${{ env.CMAKE_ARGS }}
          cmake --build build --config Release

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
-
      - name: Pack artifacts
        id: pack_artifacts
        run: |
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Redist\MSVC\14.51.36231\debug_nonredist\${{ matrix.arch }}\Microsoft.VC145.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
          7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*

      - name: Upload artifacts
@@ -748,6 +744,11 @@ jobs:
          path: llama-bin-win-cpu-${{ matrix.arch }}.zip
          name: llama-bin-win-cpu-${{ matrix.arch }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
+
  windows-rocm:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -840,11 +841,6 @@ jobs:
            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-
      - name: Verify HIP backend was built
        run: |
          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
@@ -877,6 +873,11 @@ jobs:
          path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
          name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+
  windows:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1042,11 +1043,6 @@ jobs:
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
-
      - name: Pack artifacts
        id: pack_artifacts
        run: |
@@ -1082,6 +1078,11 @@ jobs:
          path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
          name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+
  windows-sycl:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1141,11 +1142,6 @@ jobs:
            -DLLAMA_BUILD_BORINGSSL=ON
          cmake --build build --target ggml-sycl -j %NUMBER_OF_PROCESSORS%

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-sycl
-
      - name: Build the release package
        id: pack_artifacts
        run: |
@@ -1192,6 +1188,11 @@ jobs:
          path: llama-bin-win-sycl-x64.zip
          name: llama-bin-win-sycl-x64.zip

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-sycl
+
  ubuntu-24-sycl:
    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1264,11 +1265,6 @@ jobs:
            -DGGML_SYCL_F16=${{ matrix.fp16 }}
          time cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
-
      - name: Determine tag name
        id: tag
        uses: ./.github/actions/get-tag-name
@@ -1285,6 +1281,11 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
          name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz

+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
+
 #   ubuntu-22-rocm:
 #     needs: [check-release, get-version]
 #     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1379,11 +1380,6 @@ jobs:
 #             ${{ env.CMAKE_ARGS }}
 #           cmake --build build --config Release -j $(nproc)

-#       # - name: ccache-clear
-#       #   uses: ./.github/actions/ccache-clear
-#       #   with:
-#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
-
 #       - name: Determine tag name
 #         id: tag
 #         uses: ./.github/actions/get-tag-name
@@ -1403,6 +1399,11 @@ jobs:
 #           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
 #           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz

+#       # - name: ccache-clear
+#       #   uses: ./.github/actions/ccache-clear
+#       #   with:
+#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+
  ios-xcode:
    needs: [check-release, get-version]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1569,6 +1570,8 @@ jobs:
    # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
    permissions:
        contents: write # for creating release
+        id-token: write
+        attestations: write

    runs-on: ubuntu-slim

@@ -1577,14 +1580,14 @@ jobs:
      - windows
      - windows-cpu
      - windows-cuda
-      #- windows-sycl
+      - windows-sycl
      - windows-rocm
      - windows-openvino
      #- ubuntu-22-rocm
      - ubuntu-cpu
      - ubuntu-vulkan
      - ubuntu-24-openvino
-      #- ubuntu-24-sycl
+      - ubuntu-24-sycl
      - android-arm64
      - macos-cpu
      - ios-xcode
@@ -1662,6 +1665,12 @@ jobs:
        run: |
          tar -czvf release/llama-${{ steps.tag.outputs.name }}-ui.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./ui-dist .

+      - name: Attest release artifacts
+        id: attest
+        uses: actions/attest@v4
+        with:
+          subject-path: 'release/*'
+
      - name: Create and push git tag
        run: |
          TAG="${{ steps.tag.outputs.name }}"
@@ -1679,6 +1688,7 @@ jobs:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        with:
          tag_name: ${{ steps.tag.outputs.name }}
+          prerelease: true
          body: |
            <details open>

@@ -1689,6 +1699,9 @@ jobs:
            **Website:**
            - <https://llama.app>

+            **Attestations:**
+            - <${{ steps.attest.outputs.attestation-url }}>
+
            **macOS/iOS:**
            - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
            - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780)
@@ -9,6 +9,7 @@ General:

 Coding:
 - When in doubt, always refer to the CONTRIBUTING.md file of the project
+- In `test-backend-ops.cpp`, do not mention specific backends (e.g. Metal, CUDA) in comments
 - When referencing issues or PRs in comments, use the format:
  - C/C++ code: `// ref: <url>`
  - Other (CMake, etc.): `# ref: <url>`
@@ -4,8 +4,8 @@ include(CheckIncludeFileCXX)

 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 1)
-set(LLAMA_VERSION_PATCH 1)
+set(LLAMA_VERSION_MINOR 2)
+set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

 # whether this is a development/nightly build
@@ -224,10 +224,10 @@ add_subdirectory(src)
 # utils, programs, examples and tests
 #

+add_subdirectory(vendor)
+
 if (LLAMA_BUILD_COMMON)
    add_subdirectory(common)
-    add_subdirectory(vendor/cpp-httplib)
-    add_subdirectory(vendor/hash)
 endif()

 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
@@ -7,10 +7,11 @@
 <b>LLM inference in C/C++</b>

 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp?filter=v*)](https://github.com/ggml-org/llama.cpp/releases?q=tag:v0)
+[![Nightly](https://img.shields.io/github/v/release/ggml-org/llama.cpp?label=nightly)](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)
-[![Docker](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
-[![Winget](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
+[![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
+[![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)

 [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) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)

@@ -119,7 +120,7 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 ## Acknowledgements

 - [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
- [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [nothings/stb](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
 - [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
- [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
- [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
+- [mackron/miniaudio](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
+- [sheredom/subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain
@@ -290,6 +290,7 @@ combine_static_libraries() {
        "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
        "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
        "${base_dir}/${build_dir}/tools/mtmd/${release_dir}/libmtmd.a"
+        "${base_dir}/${build_dir}/vendor/hash/${release_dir}/libvendor-hash.a"
    )

    # Create temporary directory for processing
@@ -190,7 +190,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"

    # TODO: fix and re-enable the `test-llama-archs` test below
-    CTEST_EXTRA="-E test-llama-archs"
+    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
 fi

 ## helpers
@@ -8,6 +8,7 @@ set( CMAKE_CXX_COMPILER  clang++ )

 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
+set( CMAKE_ASM_COMPILER_TARGET ${target} )

 set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
@@ -126,7 +126,8 @@ set_target_properties(${TARGET} PROPERTIES
    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )

-target_include_directories(${TARGET} PUBLIC . ../vendor)
+target_include_directories(${TARGET} PUBLIC .)
+target_link_libraries     (${TARGET} PUBLIC vendor::nlohmann vendor::sheredom)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)

 if (LLAMA_SUBPROCESS)
@@ -2595,6 +2595,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.mmproj_use_gpu = value;
        }
    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
+    add_opt(common_arg(
+        // note: "-mmdev" must sort after "--rpc" in the preset map, else RPC devices are not registered yet
+        {"-mmdev", "--mmproj-device"}, "DEVICE",
+        "device to use for multimodal projector (none = don't offload, default: auto)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            if (value == "none") {
+                params.mmproj_use_gpu = false;
+                params.mmproj_device  = nullptr;
+                return;
+            }
+            auto devices = parse_device_list(value);
+            // parse_device_list pushes nullptr at back so devices is length 2 for single device.
+            if (devices.size() > 2) {
+                throw std::invalid_argument("only one device may be specified for mmproj");
+            }
+            params.mmproj_use_gpu = true;
+            params.mmproj_device  = devices.front();
+        }
+    ).set_examples(mmproj_examples).set_env("MTMD_BACKEND_DEVICE")); // no LLAMA_ARG_ prefix for backward compatibility reason
    add_opt(common_arg(
        {"--image", "--audio", "--video"}, "FILE",
        "path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files\n",
@@ -4658,6 +4678,12 @@ void common_params_add_preset_options(std::vector<common_arg> & args) {
        [](common_params &, int) { /* unused */ }
    ).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only());

+    args.push_back(common_arg(
+        {"dedup-cache-models"}, "0|1",
+        "in server router mode, hide a cached model from the model list when this preset resolves to the same model file",
+        [](common_params &, const std::string &) { /* unused */ }
+    ).set_env(COMMON_ARG_PRESET_DEDUP_CACHE_MODELS).set_preset_only());
+
    // args.push_back(common_arg(
    //     {"pin"},
    //     "in server router mode, do not unload this model if models_max is exceeded",
@@ -11,8 +11,9 @@
 #include <memory>

 // pseudo-env variable to identify preset-only arguments
-#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
-#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_LOAD_ON_STARTUP    "__PRESET_LOAD_ON_STARTUP"
+#define COMMON_ARG_PRESET_STOP_TIMEOUT       "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_DEDUP_CACHE_MODELS "__PRESET_DEDUP_CACHE_MODELS"

 //
 // CLI argument parsing
@@ -1778,6 +1778,8 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
    struct ggml_threadpool_params tpp =
            ggml_threadpool_params_from_cpu_params(params.cpuparams);

+    // each pool needs to match the respective n_threads exactly
+    // see: https://github.com/ggml-org/llama.cpp/pull/27138#issuecomment-5332307332
    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
        threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
        if (!threadpool_batch) {
@@ -581,9 +581,10 @@ struct common_params {

    // multimodal models (see tools/mtmd)
    struct common_params_model mmproj;
-    bool mmproj_use_gpu = true;     // use GPU for multimodal model
-    bool no_mmproj = false;         // explicitly disable multimodal model
-    std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
+    bool mmproj_use_gpu = true;                 // use GPU for multimodal model
+    ggml_backend_dev_t mmproj_device = nullptr; // GPU device to use for multimodal model
+    bool no_mmproj = false;                     // explicitly disable multimodal model
+    std::vector<std::string> image;             // path to image file(s) ; TODO: change the name to "media"
    int image_min_tokens = -1;
    int image_max_tokens = -1;
    int mtmd_batch_max_tokens = 1024;
@@ -989,6 +989,26 @@ std::vector<common_cached_model_info> common_list_cached_models() {
    return result;
 }

+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file) {
+    auto [repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
+
+    auto files = hf_cache::get_cached_files(repo);
+    if (files.empty()) {
+        return "";
+    }
+
+    if (!hf_file.empty()) {
+        for (const auto & f : files) {
+            if (f.path == hf_file) {
+                return f.local_path;
+            }
+        }
+        return "";
+    }
+
+    return find_best_model(files, tag).local_path;
+}
+
 bool common_download_remove(const std::string & hf_repo_with_tag) {
    namespace fs = std::filesystem;

@@ -85,6 +85,10 @@ std::vector<std::string> common_download_get_all_parts(const std::string & url);
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();

+// resolve the local cached file path for a HF repo without network access (hf_file, if given, must match exactly)
+// returns an empty string if the model is not present in the cache
+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file = "");
+
 // download single file from url to local path
 // returns status code or -1 on error
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
@@ -278,7 +278,9 @@ static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };

-static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+static const int MAX_PATTERN_DEPTH = 100;
+
+static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?', '^', '$'};
 static std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};

 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
@@ -309,6 +311,32 @@ static std::string format_literal(const std::string & literal) {

 std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }

+static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
+    if (pos + 1 >= pattern.length() || pattern[pos] != '\\') {
+        return 0;
+    }
+    size_t n_hex = 0;
+    switch (pattern[pos + 1]) {
+        case 'x': n_hex = 2; break;
+        case 'u': n_hex = 4; break;
+        case 'U': n_hex = 8; break;
+        case 't': case 'r': case 'n': case '\\': case '"': case '[': case ']':
+            return 2;
+        default:
+            return 0;
+    }
+    if (pos + 2 + n_hex > pattern.length()) {
+        return 0;
+    }
+    for (size_t i = pos + 2; i < pos + 2 + n_hex; i++) {
+        char h = pattern[i];
+        if (!((h >= '0' && h <= '9') || (h >= 'a' && h <= 'f') || (h >= 'A' && h <= 'F'))) {
+            return 0;
+        }
+    }
+    return 2 + n_hex;
+}
+
 class common_schema_converter {
 private:
    friend class common_schema_info;
@@ -345,16 +373,42 @@ private:
        return string_join(rules, " | ");
    }

+    // thrown when the pattern is a valid regex with no grammar equivalent
+    struct unsupported_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
+    // thrown when the pattern is not a valid regex
+    struct invalid_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
    std::string _visit_pattern(const std::string & pattern, const std::string & name) {
-        if (!(pattern.front() == '^' && pattern.back() == '$')) {
-            _errors.push_back("Pattern must start with '^' and end with '$'");
+        auto rules_snapshot = _rules;
+        try {
+            return _pattern_to_rule(pattern, name);
+        } catch (const unsupported_pattern & err) {
+            // revert rules
+            _rules = std::move(rules_snapshot);
+            _warnings.push_back("pattern " + pattern + " is not supported (" + err.what() + "), accepting any string");
+            return _add_rule(name, _add_primitive("string", PRIMITIVE_RULES.at("string")));
+        } catch (const invalid_pattern & err) {
+            _rules = std::move(rules_snapshot);
+            _errors.push_back("Invalid pattern " + pattern + ": " + err.what());
            return "";
        }
+    }
+
+    std::string _pattern_to_rule(const std::string & pattern, const std::string & name) {
+        if (pattern.length() < 2 || pattern.front() != '^' || pattern.back() != '$') {
+            throw unsupported_pattern("not anchored with '^' and '$'");
+        }
        std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
        std::unordered_map<std::string, std::string> sub_rule_ids;

        size_t i = 0;
        size_t length = sub_pattern.length();
+        int paren_depth = 0;

        using literal_or_rule = std::pair<std::string, bool>;
        auto to_rule = [&](const literal_or_rule & ls) {
@@ -363,7 +417,6 @@ private:
            return is_literal ? "\"" + s + "\"" : s;
        };
        std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
-            size_t start = i;
            std::vector<literal_or_rule> seq;

            auto get_dot = [&]() {
@@ -420,43 +473,42 @@ private:
                        if (i + 1 < length && sub_pattern[i + 1] == ':') {
                            i += 2; // skip "?:" for non-capturing group, treat as regular group
                        } else {
-                            // lookahead/lookbehind (?=, ?!, ?<=, ?<!) - not supported
-                            _warnings.push_back("Unsupported pattern syntax");
-                            // skip to matching ')' to avoid UB on empty seq
-                            int depth = 1;
-                            while (i < length && depth > 0) {
-                                if (sub_pattern[i] == '\\' && i + 1 < length) {
-                                    i += 2; // skip escaped character
-                                } else {
-                                    if (sub_pattern[i] == '(') depth++;
-                                    else if (sub_pattern[i] == ')') depth--;
-                                    i++;
-                                }
-                            }
-                            continue;
+                            // lookaround, named group, inline flags, ...
+                            throw unsupported_pattern("unsupported group syntax");
                        }
                    }
+                    paren_depth++;
+                    if (paren_depth > MAX_PATTERN_DEPTH) {
+                        throw unsupported_pattern("pattern nesting too deep");
+                    }
                    seq.emplace_back("(" + to_rule(transform()) + ")", false);
                } else if (c == ')') {
                    i++;
-                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
-                        _errors.push_back("Unbalanced parentheses");
+                    if (paren_depth == 0) {
+                        throw invalid_pattern("unbalanced parentheses");
                    }
+                    paren_depth--;
                    return join_seq();
+                } else if (c == '^' || c == '$') {
+                    throw unsupported_pattern("anchor inside the pattern");
                } else if (c == '[') {
                    std::string square_brackets = std::string(1, c);
                    i++;
                    while (i < length && sub_pattern[i] != ']') {
                        if (sub_pattern[i] == '\\') {
-                            square_brackets += sub_pattern.substr(i, 2);
-                            i += 2;
+                            auto escape_length = gbnf_escape_length(sub_pattern, i);
+                            if (escape_length == 0) {
+                                throw unsupported_pattern("unsupported escape in character class: " + sub_pattern.substr(i, 2));
+                            }
+                            square_brackets += sub_pattern.substr(i, escape_length);
+                            i += escape_length;
                        } else {
                            square_brackets += sub_pattern[i];
                            i++;
                        }
                    }
                    if (i >= length) {
-                        _errors.push_back("Unbalanced square brackets");
+                        throw invalid_pattern("unterminated character class");
                    }
                    square_brackets += ']';
                    i++;
@@ -465,6 +517,9 @@ private:
                    seq.emplace_back("|", false);
                    i++;
                } else if (c == '*' || c == '+' || c == '?') {
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
+                    }
                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
                    i++;
                } else if (c == '{') {
@@ -475,18 +530,19 @@ private:
                        i++;
                    }
                    if (i >= length) {
-                        _errors.push_back("Unbalanced curly brackets");
+                        throw unsupported_pattern("unterminated curly brackets");
                    }
                    curly_brackets += '}';
                    i++;
                    auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
                    int min_times = 0;
                    int max_times = std::numeric_limits<int>::max();
+                    if (nums.size() != 1 && nums.size() != 2) {
+                        throw unsupported_pattern("wrong number of values in curly brackets");
+                    }
                    try {
                        if (nums.size() == 1) {
                            min_times = max_times = std::stoi(nums[0]);
-                        } else if (nums.size() != 2) {
-                            _errors.push_back("Wrong number of values in curly brackets");
                        } else {
                            if (!nums[0].empty()) {
                                min_times = std::stoi(nums[0]);
@@ -495,9 +551,11 @@ private:
                                max_times = std::stoi(nums[1]);
                            }
                        }
-                    } catch (const std::invalid_argument & e) {
-                        _errors.push_back("Invalid number in curly brackets");
-                        return std::make_pair("", false);
+                    } catch (const std::logic_error &) {
+                        throw unsupported_pattern("invalid number in curly brackets");
+                    }
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
                    }
                    auto &last = seq.back();
                    auto &sub = last.first;
@@ -523,15 +581,22 @@ private:
                        return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
                    };
                    while (i < length) {
-                        if (sub_pattern[i] == '\\' && i < length - 1) {
+                        if (sub_pattern[i] == '\\') {
+                            if (i == length - 1) {
+                                throw invalid_pattern("trailing backslash");
+                            }
                            char next = sub_pattern[i + 1];
                            if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
                                i++;
                                literal += sub_pattern[i];
                                i++;
                            } else {
-                                literal += sub_pattern.substr(i, 2);
-                                i += 2;
+                                auto escape_length = gbnf_escape_length(sub_pattern, i);
+                                if (escape_length == 0) {
+                                    throw unsupported_pattern("unsupported escape: " + sub_pattern.substr(i, 2));
+                                }
+                                literal += sub_pattern.substr(i, escape_length);
+                                i += escape_length;
                            }
                        } else if (sub_pattern[i] == '"') {
                            literal += "\\\"";
@@ -544,14 +609,21 @@ private:
                            break;
                        }
                    }
-                    if (!literal.empty()) {
-                        seq.emplace_back(literal, true);
+                    if (literal.empty()) { // nothing was consumed, ex. a stray ']' or '}'
+                        throw unsupported_pattern(std::string("unsupported character: ") + c);
                    }
+                    seq.emplace_back(literal, true);
                }
            }
            return join_seq();
        };
-        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\"");
+
+        auto rule = to_rule(transform());
+        if (paren_depth != 0) {
+            throw invalid_pattern("unbalanced parentheses");
+        }
+
+        return _add_rule(name, "\"\\\"\" (" + rule + ") \"\\\"\"");
    }

    /*
@@ -2649,6 +2649,10 @@ void common_speculative_draft(common_speculative * spec) {
        for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) dparams.size(); ++seq_id) {
            auto & dp = dparams[seq_id];

+            if (!dp.drafting) {
+                continue;
+            }
+
            auto & result = *dp.result;

            // a new draft has been sampled
@@ -57,12 +57,16 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Qwen3DSparkModel": "qwen",
    "DSparkDraftModel": "qwen",
    "DSparkSpeculator": "qwen",
+    "Lfm2DSparkDraftModel": "qwen",
    "DeepseekV4ForCausalLM": "deepseek",
    "DeepseekV4DSparkModel": "deepseek",
    "DistilBertForMaskedLM": "bert",
    "DistilBertForSequenceClassification": "bert",
    "DistilBertModel": "bert",
    "Dots1ForCausalLM": "dots1",
+    "Dots3NoteForCausalLM": "dots3",
+    "Dots3NoteForConditionalGeneration": "dots3",
+    "Dots3NoteTextForCausalLM": "dots3",
    "DotsOCRForCausalLM": "qwen",
    "DreamModel": "dream",
    "Ernie4_5ForCausalLM": "ernie",
@@ -109,6 +113,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
+    "GraniteSWAForCausalLM": "granite",
+    "GraniteMoeSWAForCausalLM": "granite",
    "Grok1ForCausalLM": "grok",
    "GrokForCausalLM": "grok",
    "GroveMoeForCausalLM": "grovemoe",
@@ -0,0 +1,195 @@
+from __future__ import annotations
+
+import math
+import re
+
+from typing import TYPE_CHECKING, Callable, Iterable
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, gguf
+
+from .deepseek import DeepseekV2Model
+
+
+@ModelBase.register("Dots3NoteForCausalLM", "Dots3NoteForConditionalGeneration", "Dots3NoteTextForCausalLM")
+class Dots3NoteModel(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.DOTS3NOTE
+    skip_mtp = False
+    supports_mtp_export = True
+
+    # trunk layer count, stashed before indexing for filter_tensors (mirrors DeepseekV32Model)
+    _n_main_layers: int | None = None
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        hparams = self.hparams
+
+        # config file doesn't specify MTP block, detect it from model weight
+        self.n_nextn = 1 if "model.mtp.embed_tokens.weight" in self.model_tensors else 0
+        if self.n_nextn:
+            self.block_count += self.n_nextn
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        self.layer_types = hparams["layer_types"]
+        if len(self.layer_types) < hparams["num_hidden_layers"]:
+            raise ValueError("layer_types is shorter than num_hidden_layers")
+
+        if hparams.get("use_dsa", True) is not True:
+            raise ValueError("dots3-note conversion requires use_dsa=true")
+        if hparams.get("normalization", "RMSNorm") != "RMSNorm" or hparams.get("final_norm", "RMSNorm") != "RMSNorm":
+            raise ValueError("dots3-note conversion only supports RMSNorm")
+        if hparams.get("k_rope_only_layernorm", True) is not True:
+            raise ValueError("dots3-note conversion requires k_rope_only_layernorm=true")
+        if hparams.get("topk_method", "noaux_tc") != "noaux_tc" or hparams.get("scoring_func") != "sigmoid":
+            raise ValueError("dots3-note conversion only supports noaux_tc/sigmoid expert gating")
+        if hparams.get("n_group", 1) != 1 or hparams.get("topk_group", 1) != 1:
+            raise ValueError("dots3-note conversion does not support grouped expert routing")
+        if hparams.get("use_dynamic_rsf", False) or hparams.get("moe_gating_fp32", False):
+            raise ValueError("dots3-note conversion does not support use_dynamic_rsf/moe_gating_fp32")
+        for key in ("attention_gate_type", "swa_attention_gate_type"):
+            if hparams.get(key, "headwise") != "headwise":
+                raise ValueError(f"dots3-note conversion only supports headwise attention gate, got {key}={hparams.get(key)!r}")
+        if hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"] != hparams.get("swa_head_dim", 256):
+            raise ValueError("swa_head_dim must equal swa_qk_nope_head_dim + swa_qk_rope_head_dim")
+        if hparams["swa_qk_rope_head_dim"] != hparams["qk_rope_head_dim"]:
+            # both layer kinds share a single rope_dimension_count
+            raise ValueError("swa_qk_rope_head_dim must match qk_rope_head_dim")
+
+        self.apply_lora_rescale = hparams.get("apply_mla_qkv_lora_rescale", False)
+
+    def _is_swa_layer(self, bid: int) -> bool:
+        if bid >= self.hparams["num_hidden_layers"]:
+            # note: the NextN/MTP block uses the sliding-attention MLA
+            return True
+        return self.layer_types[bid] == "sliding_attention"
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endofassistant|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+        if name.startswith(("vision_encoder.", "audio_encoder.")):
+            return None
+
+        assert cls._n_main_layers is not None
+        is_mtp = name.startswith("model.mtp.") or \
+            ((m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers)
+
+        # --no-mtp: drop the NextN/MTP block; --mtp: keep only that block plus the shared embeddings/norm/lm_head
+        if is_mtp and cls.no_mtp:
+            return None
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        # head_count is a per-layer array because the two layer kinds have different head counts
+        n_layer = hparams["num_hidden_layers"]
+        hparams["num_attention_heads"] = [
+            hparams["swa_num_attention_heads"] if self._is_swa_layer(il) else hparams["num_attention_heads"]
+            for il in range(self.block_count)
+        ]
+
+        # prevent the base class from emitting key/value_length from the unused head_dim
+        hparams.pop("head_dim", None)
+
+        super().set_gguf_parameters()
+
+        # MLA geometry of the sliding-window layers (rope.freq_base_swa is emitted by the base class)
+        swa_kv_lora_rank = hparams["swa_kv_lora_rank"]
+        self.gguf_writer.add_sliding_window(hparams["sliding_window_size"])
+        self.gguf_writer.add_sliding_window_pattern([self._is_swa_layer(il) for il in range(n_layer)])
+        self.gguf_writer.add_kv_lora_rank_swa(swa_kv_lora_rank)
+        self.gguf_writer.add_key_length_swa(swa_kv_lora_rank + hparams["swa_qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_swa(swa_kv_lora_rank)
+        self.gguf_writer.add_key_length_mla_swa(hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_mla_swa(hparams["swa_v_head_dim"])
+        if hparams["swa_q_lora_rank"] != hparams["q_lora_rank"]:
+            raise ValueError("dots3-note conversion assumes a shared q_lora_rank for both layer kinds")
+
+        if self.n_nextn:
+            self.gguf_writer.add_nextn_predict_layers(self.n_nextn)
+
+        # DSA indexer (full-attention layers only)
+        self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])
+        self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])
+        self.gguf_writer.add_indexer_top_k(hparams["index_topk"])
+        self.gguf_writer.add_indexer_types([not self._is_swa_layer(il) for il in range(n_layer)])
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # move the MTP token embedding into the NextN block so the standard nextn mapping picks it up
+        if name == "model.mtp.embed_tokens.weight":
+            name = f"model.layers.{self.hparams['num_hidden_layers']}.embed_tokens.weight"
+            bid = self.hparams["num_hidden_layers"]
+
+        # fold the activation rescale sqrt(n_embd/lora_rank) into the preceding RMSNorm weight
+        # this also covers the indexer wq_b, which reads the same rescaled q_lora activation
+        if self.apply_lora_rescale and bid is not None:
+            if name.endswith("q_a_layernorm.weight"):
+                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / self.hparams["q_lora_rank"])
+            elif name.endswith("kv_a_layernorm.weight"):
+                rank = self.hparams["swa_kv_lora_rank"] if self._is_swa_layer(bid) else self.hparams["kv_lora_rank"]
+                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / rank)
+
+        # MLA absorption: split kv_b_proj into k_b (transposed) and v_b, per-layer-kind geometry
+        if name.endswith("kv_b_proj.weight"):
+            assert bid is not None
+            if self._is_swa_layer(bid):
+                n_head = self.hparams["swa_num_attention_heads"]
+                qk_nope_head_dim = self.hparams["swa_qk_nope_head_dim"]
+                v_head_dim = self.hparams["swa_v_head_dim"]
+            else:
+                n_head = self.hparams["num_attention_heads"]
+                qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+                v_head_dim = self.hparams["v_head_dim"]
+            if isinstance(n_head, list):  # set_gguf_parameters turns this into a per-layer array
+                n_head = n_head[bid]
+
+            assert data_torch.shape[0] == n_head * (qk_nope_head_dim + v_head_dim)
+
+            kv_b = data_torch.view(n_head, qk_nope_head_dim + v_head_dim, data_torch.shape[-1])
+            k_b, v_b = kv_b.split([qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+
+            yield from ModelBase.modify_tensors(self, k_b, name.replace("kv_b_proj", "k_b_proj"), bid)
+            yield from ModelBase.modify_tensors(self, v_b, name.replace("kv_b_proj", "v_b_proj"), bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -74,6 +74,108 @@ class GraniteModel(LlamaModel):
        return super().filter_tensors(item)


+@ModelBase.register("GraniteSWAForCausalLM")
+class GraniteSWAModel(GraniteModel):
+    """Conversion for IBM's GraniteSWAForCausalLM (interleaved sliding window attention)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if name.endswith("sinks"):
+            name += ".weight"
+
+        return super().filter_tensors((name, gen))
+
+    def set_gguf_parameters(self):
+        """GraniteSWA uses Granite parameters plus sliding window configuration."""
+        super().set_gguf_parameters()
+
+        # Add sliding_window from config
+        sliding_window = self.hparams.get("sliding_window", 128)
+        self.gguf_writer.add_sliding_window(sliding_window)
+        logger.info("gguf: (granite_swa) sliding_window = %s", sliding_window)
+
+        # Derive sliding_window_pattern from layer_types
+        if layer_types := self.hparams.get("layer_types"):
+            is_swa = [t == "sliding_attention" for t in layer_types]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern = %d SWA layers / %d total",
+                        sum(is_swa), len(is_swa))
+        else:
+            # Fall back to period-based pattern: i % 4 != 0
+            # This matches the transformers default pattern
+            n_layers = self.block_count
+            is_swa = [i % 4 != 0 for i in range(n_layers)]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern (inferred) = %d SWA layers / %d total",
+                        sum(is_swa), n_layers)
+
+        # Add rope_pattern from no_rope_layers
+        if no_rope_layers := self.hparams.get("no_rope_layers"):
+            # Convert 1/0 to bool (1 = use RoPE, 0 = NoPE)
+            rope_pattern = [bool(x) for x in no_rope_layers]
+            self.gguf_writer.add_rope_pattern(rope_pattern)
+            logger.info("gguf: (granite_swa) rope_pattern = %d RoPE layers / %d total",
+                        sum(rope_pattern), len(rope_pattern))
+
+
+@ModelBase.register("GraniteMoeSWAForCausalLM")
+class GraniteMoeSWAModel(GraniteSWAModel):
+    """Conversion for IBM's GraniteMoeSWAForCausalLM (unified dense + MoE with iSWA)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if shared_intermediate_size := self.hparams.get("shared_intermediate_size"):
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_intermediate_size)
+            logger.info("gguf: (granitemoewa) shared_intermediate_size = %s", shared_intermediate_size)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        """Split merged MoE tensors (gate+up) following standard MoE pattern."""
+
+        # Handle expert FFN tensors (merged gate+up) - swash format: experts.gate_up_proj
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.experts.gate_up_proj"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, f"Merged FFN tensor size must be 2 * intermediate_size, got {data_torch.shape[-2]}"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle expert FFN down projection - swash format: experts.down_proj
+        if name.endswith("block_sparse_moe.experts.down_proj"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_EXP, bid), bid)
+            return
+
+        # Handle expert FFN tensors (merged gate+up) - standard granite format: input_linear.weight
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.input_linear.weight"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle shared expert FFN tensors (if present) - kept fused since
+        # inference (build_ffn) supports a single ffn_up_shexp tensor with
+        # LLM_FFN_SWIGLU for the shared expert.
+        if name.endswith("shared_mlp.input_linear.weight"):
+            ffn_dim = self.hparams.get("shared_intermediate_size", self.hparams["intermediate_size"])
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid)
+            return
+
+        # Handle shared expert output (if present)
+        if name.endswith("shared_mlp.output_linear.weight"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, bid), bid)
+            return
+
+        # Pass through to parent for all other tensors (including sinks)
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
@ModelBase.example("ibm-granite/granite-3.1-3b-a800m-instruct")
 class GraniteMoeModel(GraniteModel):
@@ -207,7 +207,9 @@ class NemotronHModel(GraniteHybridModel):
        # calling the parent __init__. This is because the parent constructor
        # uses self.model_arch to build the tensor name map, and all MoE-specific
        # mappings would be missed if it were called with the default non-MoE arch.
-        hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
        has_moe_params = (
            "num_experts_per_tok" in hparams
            or (isinstance(hparams.get("llm_config"), dict) and "num_experts_per_tok" in hparams["llm_config"])
@@ -215,8 +217,11 @@ class NemotronHModel(GraniteHybridModel):
        if has_moe_params:
            self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
            self.is_moe = True
+        layers_block_type = hparams.get("layers_block_type")
+        if layers_block_type is not None:
+            hparams["num_hidden_layers"] = len(layers_block_type)

-        super().__init__(*args, **kwargs)
+        super().__init__(*args, hparams=hparams, **kwargs)

        # Save the top-level head_dim for later
        self.head_dim = self.hparams.get("head_dim", self.hparams.get("attention_head_dim"))
@@ -709,7 +709,7 @@ class DFlashModel(Qwen3Model):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator")
+@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator", "Lfm2DSparkDraftModel")
@ModelBase.example("satgeze/Qwen3.6-27B-DSpark")
 class DSparkModel(DFlashModel):
    # DSpark = DFlash + a semi-autoregressive Markov head.
@@ -759,6 +759,13 @@ class DSparkModel(DFlashModel):
            return None
        return super().filter_tensors(item)

+    _ROPE_PERMUTE_SUFFIXES = (
+        "self_attn.q_proj.weight",
+        "self_attn.k_proj.weight",
+        "self_attn.q_norm.weight",
+        "self_attn.k_norm.weight",
+    )
+
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if name == "model.d2t":
            self._d2t = data_torch
@@ -767,6 +774,12 @@ class DSparkModel(DFlashModel):
        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith(("embed_tokens.weight", "lm_head.weight")):
            return

+        # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd
+        if not self.hparams.get("rope_is_neox_style", True) and name.endswith(self._ROPE_PERMUTE_SUFFIXES):
+            head_dim = self.hparams["head_dim"]
+            shape = data_torch.shape
+            data_torch = data_torch.reshape(-1, head_dim // 2, 2, *shape[1:]).transpose(1, 2).reshape(shape)
+
        yield from super().modify_tensors(data_torch, name, bid)

    def prepare_tensors(self):
@@ -116,7 +116,7 @@ in inline assembler.
 Most kernels are very naive with lots of low hanging fruits left:

 > [!IMPORTANT]
-> Several assembly instructions emmited by the compiler are not implemented
+> Several assembly instructions emitted by the compiler are not implemented
 > in hardware and software emulation in firmware is not ready yet.
 > Eventually firmware will transparently trap unimplemented instructions
 > and will emulate them inside exception handler. Until then, kernel
@@ -138,12 +138,12 @@ Most kernels are very naive with lots of low hanging fruits left:
 > kernel build process. Feel free to take ideas/code from there or try linking
 > it in.

-Before commiting any changes to operations and/or kernels, don't forget
+Before committing any changes to operations and/or kernels, don't forget
 to update supported ops reports (instructions at `docs/ops.md`).

 When logging is enabled (e.g. by setting `--log-file` cli param),
 each compute kernel run outputs a line with
-pipe-delimited key-value pairs containing kernel level performance infomation.
+pipe-delimited key-value pairs containing kernel level performance information.
 Line is prefixed with `ET_PERF`:

 ```
@@ -160,7 +160,7 @@ to `GGML_ET_PROFILE/et_runtime_trace.json` and `GGML_ET_PROFILE/kernel_map` on e

 ### Uberkernel

-The in-knernel implementaiton of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
+The in-kernel implementation of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
 dispatches multiple already existing kernel implementations with device side synchronization. Due to the processor's design, there is no natural memory visibility
 horizon between sub-kernel invocations. This makes uberkernel much more difficult to develop and debug. Currently Uberkerel is hidden begind the
 `GGML_ET_UBERKERNEL` environment variable and is disabled by default. Setting it to 1 enables it and provides significant performance improvements but is only
@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail

-OPENVINO_VERSION_MAJOR="2026.2.1"
-OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"
+OPENVINO_VERSION_MAJOR="2026.3"
+OPENVINO_VERSION_FULL="2026.3.0.22451.bd8d6542e3c"

 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -334,7 +334,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.

 </details>

@@ -364,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================

-set "OPENVINO_VERSION_MAJOR=2026.2.1"
-set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"
+set "OPENVINO_VERSION_MAJOR=2026.3"
+set "OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c"

 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -547,7 +547,7 @@ endlocal
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.

 </details>

@@ -70,17 +70,23 @@ cmake --build build --config Release
    - Tab Workload: Desktop-development with C++
    - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
    - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
-    - For Windows on ARM (arm64, WoA) build with:
-    ```bash
-    cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
-    cmake --build build-arm64-windows-llvm-release
-    ```
-    For building with ninja generator and clang compiler as default:
-      -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
+    - For Windows on ARM (arm64, WoA), build with:
      ```bash
-      cmake --preset x64-windows-llvm-release
-      cmake --build build-x64-windows-llvm-release
+      cmake --preset arm64-windows-llvm-release -D GGML_OPENMP_FETCH=ON
+      cmake --build build-arm64-windows-llvm-release
      ```
+      - Use `ARM64 Native Tools Command Prompt for VS 2022` if you are building on an ARM64 machine.
+      - `GGML_OPENMP_FETCH` downloads the official LLVM OpenMP runtime and requires Clang, 7-Zip and network access during configuration. CMake selects the runtime from the target architecture, so this also works when cross-compiling for WoA from x64. The extracted header, import library, DLL and OpenMP license are placed under `build/_deps`. The build copies `libomp.dll` and `LICENSE-LLVM-OpenMP` to the runtime output directory and installs them together. Omit the option to use CMake's normal OpenMP detection, or pass `-D GGML_OPENMP=OFF` to disable OpenMP.
+    - For building with ninja generator and clang compiler as default:
+      - Set path:
+        ```
+        set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
+        ```
+      - Run:
+        ```bash
+        cmake --preset x64-windows-llvm-release
+        cmake --build build-x64-windows-llvm-release
+        ```
 - If you want HTTPS/TLS features, you may install OpenSSL development libraries. If not installed, the project will build and run without SSL support.
  - **Debian / Ubuntu:** `sudo apt-get install libssl-dev`
  - **Fedora / RHEL / Rocky / Alma:** `sudo dnf install openssl-devel`
@@ -166,6 +166,19 @@ Examples:
 - Some models require scaling the input position. For example, `[0, 1, 2, ...]` becomes `[0, 0.5, 1, ...]`. In this case, you can provide the scaling via `freq_scale = 0.5f`.
 - Some models use learned RoPE frequencies instead of relying on `powf(freq_base, -2.0 * i / n_dims)`. In this case, you can provide the learned frequencies via the `rope_freqs` tensor (corresponding to the `c` argument in `ggml_rope_ext`), then set `freq_base = 1.0f`. An important note is that `rope_freqs` in GGML is the **inverse** (`theta = pos[i] / rope_freqs`), so you may need to invert `rope_freqs` during conversion.

+### Rotating only a part of the head
+
+Many models rotate only a part of each head and leave the rest untouched (often called the "nope" part). Do not build this with views plus `ggml_concat`, it's not efficient. Both layouts can be done with a single RoPE op:
+
+- `[rope|nope]`, rotated dims first: pass `n_dims` smaller than the head size to `ggml_rope_ext`. Dims from `n_dims` to the end are copied as-is.
+- `[nope|rope]`, rotated dims last: call `ggml_rope_set_offset(cur, n_offs)` on the result of the RoPE, where `n_offs` is the size of the leading untouched part. Dims outside `[n_offs, n_offs + n_dims)` are copied as-is.
+
+`n_offs` must be even, `n_offs + n_dims` must fit in the row, and vision RoPE is not supported. Note that the frequencies are computed relative to the rotated window.
+
+Example: DeepSeek-V4 uses `[nope|rope]` for its query, key and compressed KV tensors, so `src/models/deepseek4.cpp` ropes the whole tensor and then calls `ggml_rope_set_offset(cur, n_embd_head_nope)`.
+
+Exception: some models apply an extra op to the `nope` part, for example `deepseek32.cpp`, and may not use this optimization. While RoPE can be applied selectively to a part of the head, the extra op may not, so these models still need views plus `ggml_concat`.
+
 ## GGUF specification

 https://github.com/ggml-org/ggml/blob/master/docs/gguf.md
@@ -2,5 +2,5 @@ set(TARGET llama-gguf-hash)
 add_executable(${TARGET} gguf-hash.cpp)
 install(TARGETS ${TARGET} RUNTIME)

-target_link_libraries(${TARGET} PRIVATE vendor-hash ggml ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE vendor::hash ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
@@ -17,14 +17,17 @@
 extern "C" {
 #endif

-#include "xxhash/xxhash.h"
-#include "sha1/sha1.h"
-#include "sha256/sha256.h"
+#include "hash/xxhash/xxhash.h"
+#include "hash/sha256/sha256.h"

 #ifdef __cplusplus
 }
 #endif

+// sha1 is compiled as C++ and lives in a namespace, see scripts/sync_vendor.py
+#include "hash/sha1/sha1.h"
+using namespace vendor_hash;
+

 // uuid.uuid5(uuid.NAMESPACE_URL, 'en.wikipedia.org/wiki/Llama.cpp')
 #define UUID_NAMESPACE_LLAMA_CPP "ef001206-dadc-5f6d-a15f-3359e577d4e5"
@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)

 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 20)
-set(GGML_VERSION_PATCH 1)
+set(GGML_VERSION_MINOR 21)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -243,6 +243,7 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                            "ggml: metal minimum macOS version")
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
+option(GGML_OPENMP_FETCH                    "ggml: fetch LLVM OpenMP"                         OFF)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
@@ -341,9 +342,6 @@ set(GGML_PUBLIC_HEADERS
    include/gguf.h)

 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-#if (GGML_METAL)
-#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
-#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)

@@ -7,7 +7,7 @@ extern "C" {
 #endif

 #define RPC_PROTO_MAJOR_VERSION    5
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MINOR_VERSION    1
 #define RPC_PROTO_PATCH_VERSION    0

 #ifdef  __cplusplus
@@ -1981,6 +1981,14 @@ extern "C" {
            float                 beta_fast,
            float                 beta_slow);

+    // set the offset dims for RoPE
+    // a must be GGML_OP_ROPE or GGML_OP_ROPE_BACK
+    // vision RoPE is not supported
+    // example: (marking: x = rotated, 0 = unrotated)
+    //     n_embd = 10, n_dims = 4, offset = 2 --> [00xxxx0000]
+    GGML_API struct ggml_tensor * ggml_rope_set_offset(
+            struct ggml_tensor  * a,
+            int                   n_offs);

    // clamp
    // in-place, returns view(a)
@@ -222,9 +222,123 @@ if (GGML_SCHED_NO_REALLOC)
    target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC)
 endif()

-if (GGML_OPENMP)
+if (GGML_OPENMP_FETCH)
+    if (NOT GGML_OPENMP)
+        message(FATAL_ERROR "GGML_OPENMP_FETCH requires GGML_OPENMP")
+    elseif (NOT WIN32 OR NOT (CMAKE_C_COMPILER_ID MATCHES "Clang"))
+        message(FATAL_ERROR "GGML_OPENMP_FETCH currently requires Clang on Windows")
+    endif()
+
+    set(GGML_OPENMP_LLVM_VERSION "20.1.8")
+    string(REGEX MATCH "^[0-9]+" GGML_OPENMP_LLVM_VERSION_MAJOR "${GGML_OPENMP_LLVM_VERSION}")
+    string(REGEX MATCH "^[0-9]+" GGML_OPENMP_COMPILER_VERSION_MAJOR "${CMAKE_C_COMPILER_VERSION}")
+    if (NOT GGML_OPENMP_COMPILER_VERSION_MAJOR STREQUAL GGML_OPENMP_LLVM_VERSION_MAJOR)
+        message(FATAL_ERROR "LLVM OpenMP ${GGML_OPENMP_LLVM_VERSION} requires Clang ${GGML_OPENMP_LLVM_VERSION_MAJOR}.x")
+    endif()
+
+    string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" GGML_OPENMP_SYSTEM_PROCESSOR)
+    if (GGML_OPENMP_SYSTEM_PROCESSOR MATCHES "^(amd64|x86_64)$")
+        set(GGML_OPENMP_ARCH "x64")
+        set(GGML_OPENMP_INSTALLER_SUFFIX "win64")
+        set(GGML_OPENMP_INSTALLER_SHA256 "3197846a2b19063687dd56e93e34cd941e3548d907f23a6131571321bdf9fe7b")
+    elseif (GGML_OPENMP_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm64)$")
+        set(GGML_OPENMP_ARCH "arm64")
+        set(GGML_OPENMP_INSTALLER_SUFFIX "woa64")
+        set(GGML_OPENMP_INSTALLER_SHA256 "7c4ac97eb2ae6b960ca5f9caf3ff6124c8d2a18cc07a7840a4d2ea15537bad8e")
+    else()
+        message(FATAL_ERROR "GGML_OPENMP_FETCH does not support ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+
+    set(GGML_OPENMP_CACHE_DIR "${CMAKE_BINARY_DIR}/_deps")
+    set(GGML_OPENMP_ROOT "${GGML_OPENMP_CACHE_DIR}/llvm-openmp-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_ARCH}")
+    set(GGML_OPENMP_LIBRARY "${GGML_OPENMP_ROOT}/lib/libomp.lib")
+    set(GGML_OPENMP_RUNTIME "${GGML_OPENMP_ROOT}/bin/libomp.dll")
+    set(GGML_OPENMP_HEADER "${GGML_OPENMP_ROOT}/include/omp.h")
+    set(GGML_OPENMP_LICENSE "${GGML_OPENMP_ROOT}/LICENSE.TXT")
+    set(GGML_OPENMP_LICENSE_SHA256 "fdad1758a9e1f9d5a81e18879b3406772115edc92c24bfa36b70c654f325e8e4")
+
+    if (NOT EXISTS "${GGML_OPENMP_LIBRARY}" OR NOT EXISTS "${GGML_OPENMP_RUNTIME}" OR NOT EXISTS "${GGML_OPENMP_HEADER}")
+        find_program(GGML_OPENMP_7Z NAMES 7z 7zz 7za)
+        if (NOT GGML_OPENMP_7Z)
+            message(FATAL_ERROR "GGML_OPENMP_FETCH requires 7-Zip to extract the LLVM installer")
+        endif()
+
+        set(GGML_OPENMP_INSTALLER "${GGML_OPENMP_ROOT}/LLVM-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_INSTALLER_SUFFIX}.exe")
+        set(GGML_OPENMP_EXTRACT_DIR "${GGML_OPENMP_ROOT}/extract")
+        set(GGML_OPENMP_INSTALLER_URL "https://github.com/llvm/llvm-project/releases/download/llvmorg-${GGML_OPENMP_LLVM_VERSION}/LLVM-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_INSTALLER_SUFFIX}.exe")
+
+        file(MAKE_DIRECTORY "${GGML_OPENMP_EXTRACT_DIR}")
+        file(DOWNLOAD "${GGML_OPENMP_INSTALLER_URL}" "${GGML_OPENMP_INSTALLER}"
+            EXPECTED_HASH "SHA256=${GGML_OPENMP_INSTALLER_SHA256}"
+            SHOW_PROGRESS
+            STATUS GGML_OPENMP_DOWNLOAD_STATUS)
+        list(GET GGML_OPENMP_DOWNLOAD_STATUS 0 GGML_OPENMP_DOWNLOAD_RESULT)
+        if (NOT GGML_OPENMP_DOWNLOAD_RESULT EQUAL 0)
+            list(GET GGML_OPENMP_DOWNLOAD_STATUS 1 GGML_OPENMP_DOWNLOAD_ERROR)
+            message(FATAL_ERROR "Failed to download LLVM OpenMP: ${GGML_OPENMP_DOWNLOAD_ERROR}")
+        endif()
+
+        execute_process(
+            COMMAND "${GGML_OPENMP_7Z}" e -y "-o${GGML_OPENMP_EXTRACT_DIR}" "${GGML_OPENMP_INSTALLER}" -r libomp.lib libomp.dll omp.h
+            RESULT_VARIABLE GGML_OPENMP_EXTRACT_RESULT
+            OUTPUT_QUIET)
+        if (NOT GGML_OPENMP_EXTRACT_RESULT EQUAL 0 OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/libomp.lib" OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/libomp.dll" OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/omp.h")
+            message(FATAL_ERROR "Failed to extract libomp from ${GGML_OPENMP_INSTALLER}")
+        endif()
+
+        file(MAKE_DIRECTORY "${GGML_OPENMP_ROOT}/lib" "${GGML_OPENMP_ROOT}/bin" "${GGML_OPENMP_ROOT}/include")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/libomp.lib" DESTINATION "${GGML_OPENMP_ROOT}/lib")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/libomp.dll" DESTINATION "${GGML_OPENMP_ROOT}/bin")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/omp.h" DESTINATION "${GGML_OPENMP_ROOT}/include")
+        file(REMOVE_RECURSE "${GGML_OPENMP_INSTALLER}" "${GGML_OPENMP_EXTRACT_DIR}")
+    endif()
+
+    # The NSIS installer embeds LLVM's general license in its UI but does not install it as a file; use OpenMP's license to include its additional notices.
+    if (EXISTS "${GGML_OPENMP_LICENSE}")
+        file(SHA256 "${GGML_OPENMP_LICENSE}" GGML_OPENMP_LICENSE_ACTUAL_SHA256)
+    endif()
+    if (NOT GGML_OPENMP_LICENSE_ACTUAL_SHA256 STREQUAL GGML_OPENMP_LICENSE_SHA256)
+        file(DOWNLOAD "https://raw.githubusercontent.com/llvm/llvm-project/llvmorg-${GGML_OPENMP_LLVM_VERSION}/openmp/LICENSE.TXT" "${GGML_OPENMP_LICENSE}"
+            EXPECTED_HASH "SHA256=${GGML_OPENMP_LICENSE_SHA256}")
+    endif()
+
+    if (COMMAND license_add_file)
+        license_add_file("LLVM OpenMP" "${GGML_OPENMP_LICENSE}")
+    endif()
+
+    add_library(ggml-openmp-c INTERFACE)
+    target_compile_options(ggml-openmp-c INTERFACE "$<$<COMPILE_LANGUAGE:C>:-fopenmp=libomp>")
+    target_include_directories(ggml-openmp-c SYSTEM INTERFACE "${GGML_OPENMP_ROOT}/include")
+    target_link_libraries(ggml-openmp-c INTERFACE "${GGML_OPENMP_LIBRARY}")
+
+    add_library(ggml-openmp-cxx INTERFACE)
+    target_compile_options(ggml-openmp-cxx INTERFACE "$<$<COMPILE_LANGUAGE:CXX>:-fopenmp=libomp>")
+    target_include_directories(ggml-openmp-cxx SYSTEM INTERFACE "${GGML_OPENMP_ROOT}/include")
+    target_link_libraries(ggml-openmp-cxx INTERFACE "${GGML_OPENMP_LIBRARY}")
+
+    set(GGML_OPENMP_RUNTIME_OUTPUT_DIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
+    if (CMAKE_CONFIGURATION_TYPES)
+        string(APPEND GGML_OPENMP_RUNTIME_OUTPUT_DIR "/$<CONFIG>")
+    endif()
+    add_custom_target(ggml-openmp-runtime ALL
+        COMMAND ${CMAKE_COMMAND} -E make_directory "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}"
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different "${GGML_OPENMP_RUNTIME}" "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}/libomp.dll"
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different "${GGML_OPENMP_LICENSE}" "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}/LICENSE-LLVM-OpenMP")
+    add_dependencies(ggml-base ggml-openmp-runtime)
+    install(FILES "${GGML_OPENMP_RUNTIME}" DESTINATION ${CMAKE_INSTALL_BINDIR})
+    install(FILES "${GGML_OPENMP_LICENSE}" DESTINATION ${CMAKE_INSTALL_BINDIR} RENAME LICENSE-LLVM-OpenMP)
+
+    set(GGML_OPENMP_TARGET_C ggml-openmp-c)
+    set(GGML_OPENMP_TARGET_CXX ggml-openmp-cxx)
+    set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
+elseif (GGML_OPENMP)
    find_package(OpenMP)
    if (OpenMP_FOUND)
+        set(GGML_OPENMP_TARGET_C OpenMP::OpenMP_C)
+        set(GGML_OPENMP_TARGET_CXX OpenMP::OpenMP_CXX)
        set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
    else()
        set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
@@ -236,7 +350,7 @@ endif()

 if (GGML_OPENMP_ENABLED)
    target_compile_definitions(ggml-base PRIVATE GGML_USE_OPENMP)
-    target_link_libraries(ggml-base PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+    target_link_libraries(ggml-base PRIVATE ${GGML_OPENMP_TARGET_C} ${GGML_OPENMP_TARGET_CXX})
 endif()

 add_library(ggml
@@ -1599,11 +1599,23 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
    std::vector<int32_t> ids;
    std::vector<ggml_bitset_t> used_ids;

+    int prev_backend_id = -1;
+
    for (int split_id = 0; split_id < sched->n_splits; split_id++) {
        struct ggml_backend_sched_split * split = &splits[split_id];
        int split_backend_id = split->backend_id;
        ggml_backend_t split_backend = sched->backends[split_backend_id];

+        // ensure the previous split's async work has completed before we start
+        // this split, the allocator may have reused buffer regions across splits
+        if (split->n_inputs == 0 && prev_backend_id >= 0 && prev_backend_id != split_backend_id) {
+            if (sched->events[prev_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_synchronize(sched->events[prev_backend_id][sched->cur_copy]);
+            } else {
+                ggml_backend_synchronize(sched->backends[prev_backend_id]);
+            }
+        }
+
        // copy the input tensors to the split backend
        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
@@ -1766,12 +1778,12 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
            }
        }

-        // record the event of this copy
-        if (split->n_inputs > 0) {
-            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
-            }
+        // record the event of this split
+        if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+            ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
        }
+
+        prev_backend_id = split_backend_id;
    }

    return GGML_STATUS_SUCCESS;
@@ -2534,6 +2534,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
            }
        case GGML_OP_ROPE:
            {
+                if (((const int32_t *) op->op_params)[15] != 0) {
+                    return false; // FIXME: support ggml_rope_set_offset
+                }
                if (op->src[0]->ne[0] > 896) {
                    return false;
                }
@@ -74,7 +74,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)

    if (GGML_OPENMP_ENABLED)
        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
-        target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+        target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_OPENMP_TARGET_C} ${GGML_OPENMP_TARGET_CXX})
    endif()

    if (GGML_LLAMAFILE)
@@ -639,6 +639,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/
            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)

        set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -701,6 +702,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla_asm.S
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
@@ -737,8 +740,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        set_target_properties(${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(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
-    endif()
+  if (CMAKE_C_COMPILER_ID STREQUAL "IntelLLVM" OR CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+      # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+      target_compile_options(${GGML_CPU_NAME} PRIVATE "$<$<OR:$<COMPILE_LANG_AND_ID:C,IntelLLVM>,$<COMPILE_LANG_AND_ID:CXX,IntelLLVM>>:$<$<BOOL:${WIN32}>:/clang:>-fno-associative-math>")
+  endif()
+
 endfunction()
@@ -23,6 +23,7 @@
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
 #include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.h"
 #include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.h"

 #include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
@@ -76,6 +77,21 @@ static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
    Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
 }

+template <void (*Fn)(size_t, size_t, size_t, const void *, size_t, const void *, void *, size_t, size_t, float, float)>
+static inline void kernel_run_lhs_stride_fn10(size_t       m,
+                                              size_t       n,
+                                              size_t       k,
+                                              size_t       lhs_stride,
+                                              const void * lhs,
+                                              const void * rhs,
+                                              void *       dst,
+                                              size_t       dst_stride_row,
+                                              size_t       dst_stride_col,
+                                              float        clamp_min,
+                                              float        clamp_max) {
+    Fn(m, n, k, lhs, lhs_stride, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
+}
+
 template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
 static inline void kernel_run_float_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
                                         const void* lhs, const void* rhs, void* dst,
@@ -947,25 +963,25 @@ static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = {
            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
        },
-        /* SME GEMV */
+        /* SME2 GEMV */
        {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ nullptr,
-            /* .get_rhs_packed_offset_ex = */ nullptr,
-            /* .run_kernel_ex         = */ nullptr,
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_mr                = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .run_kernel_ex         = */ &kernel_run_lhs_stride_fn10<kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
        },
        /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_f32p2vlx1_f32_sme,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
+            /* .get_offset            = */ nullptr,
+            /* .get_packed_offset_ex  = */ nullptr,
+            /* .packed_size_ex        = */ nullptr,
+            /* .pack_func_ex          = */ nullptr,
        },
        /* .rhs_info = */ {
            /* .packed_stride         = */ nullptr,
@@ -696,6 +696,15 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        }

        if (op->src[0]->type == GGML_TYPE_F32) {
+            ggml_kleidiai_kernels * primary     = kernel_chain[0];
+            kernel_info *           gemv_kernel = primary ? &primary->gemv : nullptr;
+            if (is_gemv && op->src[1]->nb[0] == (int64_t) sizeof(float) && gemv_kernel &&
+                gemv_kernel->get_lhs_offset_ex && gemv_kernel->get_rhs_packed_offset_ex &&
+                gemv_kernel->run_kernel_ex && gemv_kernel->get_dst_offset) {
+                size = 0;
+                return true;
+            }
+
            size_t cursor = 0;
            bool any_slot = false;

@@ -811,15 +820,28 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            return false;
        }

-        kernel_info * kernel        = &kernels->gemm;
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+        const bool use_gemv = m == 1 && src1->nb[0] == (int64_t) sizeof(float) &&
+                              kernels->gemv.get_lhs_offset_ex &&
+                              kernels->gemv.get_rhs_packed_offset_ex &&
+                              kernels->gemv.run_kernel_ex &&
+                              kernels->gemv.get_dst_offset;
+
+        kernel_info * kernel        = use_gemv ? &kernels->gemv : &kernels->gemm;
        lhs_packing_info * lhs_info = &kernels->gemm_lhs_info;

-        if (!kernel || !lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
-            !lhs_info->packed_size_ex || !lhs_info->pack_func_ex ||
+        if (!kernel || !kernel->get_lhs_offset_ex ||
            !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
            return false;
        }

+        if (!use_gemv && (!lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
+                          !lhs_info->packed_size_ex || !lhs_info->pack_func_ex)) {
+            return false;
+        }
+
        const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data);
        const bool has_header = kleidiai_is_weight_header_valid(header);

@@ -832,16 +854,14 @@ class tensor_traits : public ggml::cpu::tensor_traits {
        const int nth = params->nth > 0 ? params->nth : 1;
        const int ith = params->ith;

-        const size_t k = ne00;
-        const size_t m = ne11;
-        const size_t n = ne01;
-
        const size_t mr = kernel->get_mr();
        const size_t kr = kernel->get_kr();
        const size_t sr = kernel->get_sr();

-        const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
-        GGML_ASSERT(lhs_packed_size <= params->wsize);
+        const size_t lhs_packed_size = use_gemv ? 0 : lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
+        if (!use_gemv) {
+            GGML_ASSERT(lhs_packed_size <= params->wsize);
+        }

        uint8_t * lhs_packed   = static_cast<uint8_t *>(params->wdata);
        const size_t dst_stride = dst->nb[1];
@@ -853,7 +873,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
            const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
            uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];

-            {
+            if (!use_gemv) {
                const int64_t m_roundup_mr = kai_roundup((int64_t)m, (int64_t)mr);
                int64_t max_threads = mr ? (m_roundup_mr / (int64_t)mr) : nth;
                max_threads = std::max<int64_t>(1, max_threads);
@@ -903,15 +923,17 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                const size_t n_to_process = std::min(chunk_cols, n - n_start);

                if (n_to_process > 0) {
-                    const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
+                    const size_t lhs_offset = use_gemv ? kernel->get_lhs_offset_ex(0, k, 0)
+                                                       : lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
                    const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
                    const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);

-                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
+                    const void * lhs_ptr = use_gemv ? lhs_batch_base + lhs_offset
+                                                    : lhs_packed + lhs_offset;
                    const void * rhs_ptr = rhs_base + rhs_packed_offset;
                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch_base + dst_offset);

-                    kernel->run_kernel_ex(m, n_to_process, k, 0,
+                    kernel->run_kernel_ex(m, n_to_process, k, use_gemv ? src1->nb[1] : 0,
                                          lhs_ptr,
                                          rhs_ptr,
                                          dst_ptr,
@@ -5979,6 +5979,8 @@ static void ggml_compute_forward_rope_flt(
    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);

+    const int n_offs = ((int32_t *) dst->op_params)[15];
+
    GGML_TENSOR_UNARY_OP_LOCALS

    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
@@ -5995,6 +5997,10 @@ static void ggml_compute_forward_rope_flt(
    GGML_ASSERT(n_dims <= ne0);
    GGML_ASSERT(n_dims % 2 == 0);

+    GGML_ASSERT(n_offs >= 0);
+    GGML_ASSERT(n_offs % 2 == 0);
+    GGML_ASSERT(n_offs + n_dims <= ne0);
+
    // rows per thread
    const int dr = (nr + nth - 1)/nth;

@@ -6020,6 +6026,7 @@ static void ggml_compute_forward_rope_flt(

    if (is_vision) {
        GGML_ASSERT(n_dims == ne0/2);
+        GGML_ASSERT(n_offs == 0);
    }

    const float * freq_factors = NULL;
@@ -6068,12 +6075,12 @@ static void ggml_compute_forward_rope_flt(

                switch (mode) {
                    case GGML_ROPE_TYPE_NORMAL:
-                        rotate_pairs<T>(n_dims, 1, cache, src, dst_data, 1);
+                        rotate_pairs<T>(n_dims, 1, cache, src + n_offs, dst_data + n_offs, 1);
                        break;
                    case GGML_ROPE_TYPE_NEOX:
                    case GGML_ROPE_TYPE_MROPE:
                    case GGML_ROPE_TYPE_IMROPE:
-                        rotate_pairs<T>(n_dims, n_dims/2, cache, src, dst_data);
+                        rotate_pairs<T>(n_dims, n_dims/2, cache, src + n_offs, dst_data + n_offs);
                        break;
                    case GGML_ROPE_TYPE_VISION:
                        rotate_pairs<T>(ne0, n_dims, cache, src, dst_data);
@@ -6084,7 +6091,11 @@ static void ggml_compute_forward_rope_flt(

                if (!is_vision) {
                    // fill the remain channels with data from src tensor
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+                        if (i0 == n_offs) {
+                            i0 += n_dims - 2; // skip the rotated channels
+                            continue;
+                        }
                        const T * const src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                        T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);

@@ -29,13 +29,15 @@ extern "C" {
 // FP16 to FP32 conversion

 // 16-bit float
-// on Arm, we use __fp16
+// on Arm, we use __fp16, which requires the IEEE fp16 format: implied on
+// AArch64, selected by -mfp16-format=ieee on 32 bit Arm, where the compiler
+// may otherwise reject the type
 // on x86, we use uint16_t
 //
 // for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
 // for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
 //
-#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+#if defined(__ARM_NEON) && defined(__ARM_FP16_FORMAT_IEEE) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)

@@ -326,7 +328,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
    #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
 #endif

-#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
+#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA) && defined(__ARM_FP16_FORMAT_IEEE)

 #define GGML_SIMD

@@ -1418,7 +1418,9 @@ struct ggml_backend_cuda_context {
    cudaEvent_t copy_event = nullptr;

    cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
-    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
+    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {nullptr};
+    void * cublas_workspaces[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {nullptr};
+    size_t cublas_workspace_sizes[GGML_CUDA_MAX_DEVICES] = {0};

    int curr_stream_no = 0;

@@ -1495,17 +1497,22 @@ struct ggml_backend_cuda_context {

    ggml_cuda_stream_context & stream_context() { return concurrent_stream_context; }

-    cublasHandle_t cublas_handle(int device) {
-        if (cublas_handles[device] == nullptr) {
-            ggml_cuda_set_device(device);
-            CUBLAS_CHECK(cublasCreate(&cublas_handles[device]));
-            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH));
-        }
-        return cublas_handles[device];
-    }
-
    cublasHandle_t cublas_handle() {
-        return cublas_handle(device);
+        if (cublas_handles[device][curr_stream_no] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUBLAS_CHECK(cublasCreate(&cublas_handles[device][curr_stream_no]));
+            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device][curr_stream_no], CUBLAS_TF32_TENSOR_OP_MATH));
+            CUBLAS_CHECK(cublasSetStream(cublas_handles[device][curr_stream_no], stream()));
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && (CUBLAS_VER_MAJOR > 11 || (CUBLAS_VER_MAJOR == 11 && CUBLAS_VER_MINOR >= 2))
+            if (cublas_workspace_sizes[device] == 0) {
+                const int cc = ggml_cuda_info().devices[device].cc;
+                cublas_workspace_sizes[device] = (cc >= GGML_CUDA_CC_HOPPER) ? 32 * 1024 * 1024 : 4 * 1024 * 1024;
+            }
+            CUDA_CHECK(cudaMalloc(&cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
+            CUBLAS_CHECK(cublasSetWorkspace(cublas_handles[device][curr_stream_no], cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
+#endif
+        }
+        return cublas_handles[device][curr_stream_no];
    }

    // pool
@@ -711,9 +711,12 @@ ggml_backend_cuda_context::~ggml_backend_cuda_context() {
            if (streams[i][j] != nullptr) {
                CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
            }
-        }
-        if (cublas_handles[i] != nullptr) {
-            CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+            if (cublas_handles[i][j] != nullptr) {
+                CUBLAS_CHECK(cublasDestroy(cublas_handles[i][j]));
+            }
+            if (cublas_workspaces[i][j] != nullptr) {
+                CUDA_CHECK(cudaFree(cublas_workspaces[i][j]));
+            }
        }
    }
 }
@@ -1416,7 +1419,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const

    const int64_t ne_dst = ggml_nelements(dst);
    cudaStream_t main_stream = ctx.stream();
-    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
+    cublasHandle_t cublas_h = ctx.cublas_handle();

    const size_t src0_ts = ggml_type_size(src0->type);
    GGML_ASSERT(nb00 == src0_ts);
@@ -1539,14 +1542,14 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
    //     probably because the internal kernel selection logic is suboptimal.
    if (compute_type == GGML_TYPE_F32 && ne12 == 1 && ne13 == 1) {
        CUBLAS_CHECK(
-            cublasSgemm(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+            cublasSgemm(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                    ne01, ne11, ne10,
                    (const float *) alpha, (const float *) src0_ptr, s01,
                                           (const float *) src1_ptr, s11,
                    (const float *) beta,  (float       *)  dst_ptr, ne0));
    } else if (ne12 == 1 && ne13 == 1) {
        CUBLAS_CHECK(
-            cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+            cublasGemmEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                    ne01, ne11, ne10,
                    alpha, src0_ptr, cu_data_type_a, s01,
                           src1_ptr, cu_data_type_b, s11,
@@ -1561,7 +1564,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
        // 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,
+        cublasGemmStridedBatchedEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
                alpha, src0_ptr, cu_data_type_a, s01, sma,     // strideA
                       src1_ptr, cu_data_type_b, s11, smb,     // strideB
@@ -1599,7 +1602,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
        CUDA_CHECK(cudaGetLastError());

        CUBLAS_CHECK(
-        cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+        cublasGemmBatchedEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                ne01, ne11, ne10,
                alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, s01,
                       (const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
@@ -2723,6 +2726,12 @@ static bool ggml_cuda_should_fuse_rms_norm_mul_rope(const ggml_tensor * rms_norm
        return false;
    }

+    // ggml_rope_set_offset is not yet supported in the fused kernel
+    const int n_offs = ((const int32_t *) rope->op_params)[15];
+    if (n_offs != 0) {
+        return false;
+    }
+
    return true;
 }

@@ -4,6 +4,7 @@
 #include "vecdotq.cuh"

 #include <cstdint>
+#include <type_traits>

 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);

@@ -69,7 +70,8 @@ enum mmvq_parameter_table_id {
    MMVQ_PARAMETERS_GCN,
    MMVQ_PARAMETERS_RDNA2,
    MMVQ_PARAMETERS_RDNA3_0,
-    MMVQ_PARAMETERS_RDNA4
+    MMVQ_PARAMETERS_RDNA4,
+    MMVQ_PARAMETERS_GB10
 };

 static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
@@ -83,6 +85,8 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
    return MMVQ_PARAMETERS_GCN;
 #elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING && __CUDA_ARCH__ < GGML_CUDA_CC_AMPERE
    return MMVQ_PARAMETERS_TURING;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ == GGML_CUDA_CC_DGX_SPARK
+    return MMVQ_PARAMETERS_GB10;
 #else
    return MMVQ_PARAMETERS_GENERIC;
 #endif
@@ -104,6 +108,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING && ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_AMPERE) {
        return MMVQ_PARAMETERS_TURING;
    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_DGX_SPARK) {
+        return MMVQ_PARAMETERS_GB10;
+    }
    return MMVQ_PARAMETERS_GENERIC;
 }

@@ -283,6 +290,42 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
    if (!ggml_is_quantized(type)) {
        return false;
    }
+    // k-quants cost more to decode and mvq redoes that per column, so MMQ wins sooner.
+    // Only list quant-types MMQ supports, others would fall back to cuBLAS.
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_ADA_LOVELACE) {
+        switch (type) { // tuned on RTX 4090
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 4;
+            case GGML_TYPE_Q3_K:
+                return ne11 <= 6;
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_BLACKWELL) {
+        switch (type) { // tuned on RTX 5090
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 5;
+            case GGML_TYPE_Q6_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_DGX_SPARK) {
+        switch (type) { // tuned on DGX Spark GB10
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 6;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
    if (GGML_CUDA_CC_IS_CDNA(cc)) {
        if (GGML_CUDA_CC_IS_CDNA1(cc)) {
            switch (type) {
@@ -351,7 +394,7 @@ static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
 #endif
 }

-static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id, bool small_k = false, bool halve_iters = false) {
    if (table_id == MMVQ_PARAMETERS_GENERIC) {
        switch (ncols_dst) {
            case 1:
@@ -454,11 +497,32 @@ static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_d
                return 1;
        }
    }
+    if (table_id == MMVQ_PARAMETERS_GB10) {
+        const int generic = calc_nwarps(type, ncols_dst, MMVQ_PARAMETERS_GENERIC);
+        // Only worth the wider block when it actually retires the K loop in half the trips (Observation)
+        if (ncols_dst == 1 && !small_k && halve_iters) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_IQ4_NL:
+                    return 2 * generic;
+                default:
+                    break;
+            }
+        }
+        return generic;
+    }
    return 1;
 }

 static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id, bool small_k = false, int nwarps = 1) {
-    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING) {
+    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING || table_id == MMVQ_PARAMETERS_GB10) {
        switch (ncols_dst) {
            case 1:
                return small_k ? nwarps : 1;
@@ -477,8 +541,8 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
    return 1;
 }

-template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false>
-__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false, bool halve_iters = false>
+__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id(), small_k, halve_iters)*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
        const void * vx_ptr, const void * vy_ptr, const int32_t * ids_ptr, const ggml_cuda_mm_fusion_args_device fusion, float * dst_ptr,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -495,7 +559,7 @@ static __global__ void mul_mat_vec_q(
    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
    constexpr int vdr = get_vdr_mmvq(type);
    constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id);
+    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id, small_k, halve_iters);
    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();

@@ -773,8 +837,8 @@ static __global__ void mul_mat_vec_q_moe(
 template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
-        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false) {
-    const int nwarps = calc_nwarps(type, ncols_dst, table_id);
+        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false, const bool halve_iters = false) {
+    const int nwarps = calc_nwarps(type, ncols_dst, table_id, small_k, halve_iters);
    const int rpb = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
    const int64_t nblocks = (nrows_x + rpb - 1) / rpb;
    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
@@ -782,7 +846,7 @@ static std::pair<dim3, dim3> calc_launch_params(
    return {block_nums, block_dims};
 }

-template<ggml_type type, int c_ncols_dst, bool small_k = false>
+template<ggml_type type, int c_ncols_dst, bool small_k = false, bool halve_iters = false>
 static void mul_mat_vec_q_switch_fusion(
        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -797,7 +861,7 @@ static void mul_mat_vec_q_switch_fusion(
    if constexpr (c_ncols_dst == 1) {
        if (has_fusion) {
            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
-            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k>, launch_params,
+            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k, halve_iters>, launch_params,
                 vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -808,7 +872,7 @@ static void mul_mat_vec_q_switch_fusion(
    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");

    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
-    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k>, launch_params,
+    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k, halve_iters>, launch_params,
        vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -860,16 +924,18 @@ static void mul_mat_vec_q_switch_ncols_dst(

    const bool has_ids = ids != nullptr;

+    // How the K loop divides up at the baseline block width, both decisions below use these.
+    constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr                   = get_vdr_mmvq(type);
+    const int     blocks_per_row_x      = ncols_x / qk;
+    const int     blocks_per_iter_1warp = vdr * warp_size / qi;
+
    const auto should_use_small_k = [&](int c_ncols_dst) {
        // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
        // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
-        constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
-        constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
-        constexpr int vdr                   = get_vdr_mmvq(type);
-        const int     blocks_per_row_x      = ncols_x / qk;
-        const int     blocks_per_iter_1warp = vdr * warp_size / qi;
-        const int     nwarps                = calc_nwarps(type, c_ncols_dst, table_id);
-        bool          use                   = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+        const int  nwarps = calc_nwarps(type, c_ncols_dst, table_id);
+        bool       use    = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;

        constexpr std::array<ggml_type, 2> iq_slow_turing = {
            GGML_TYPE_IQ3_XXS,
@@ -902,6 +968,28 @@ static void mul_mat_vec_q_switch_ncols_dst(
        return use;
    };

+    // Whether doubling nwarps pays off on the ncols_dst == 1 path, where K sets the K loop trip count.
+    const auto should_halve_iters = [&] {
+        if (table_id != MMVQ_PARAMETERS_GB10) {
+            return false;
+        }
+
+        // Expert rows are gathered per token, so a wider block adds reduction work without reuse.
+        if (has_ids) {
+            return false;
+        }
+
+        const int blocks_per_iter = calc_nwarps(type, 1, table_id) * blocks_per_iter_1warp;
+        const int iters           = (blocks_per_row_x + blocks_per_iter - 1) /  blocks_per_iter;
+        const int iters_wide      = (blocks_per_row_x + blocks_per_iter * 2 - 1) / (blocks_per_iter * 2);
+
+        // An odd trip count leaves half the wider block idle for its last iteration, that tail is
+        // only affordable once the loop is long enough to dilute it to an eighth of the work (observation).
+        const int idle = iters_wide * 2 - iters;
+
+        return idle * 8 <= iters_wide * 2;
+    };
+
    if (has_ids && ncols_dst > 1) {
        // Multi-token MUL_MAT_ID path - dedicated MoE kernel
        mul_mat_vec_q_moe_launch<type>(
@@ -914,26 +1002,34 @@ static void mul_mat_vec_q_switch_ncols_dst(

    switch (ncols_dst) {
        case 1: {
-            constexpr int c_ncols_dst = 1;
+            // static, else MSVC lambda capture breaks the constexpr uses below
+            static constexpr int c_ncols_dst = 1;

-            bool use_small_k = should_use_small_k(c_ncols_dst);
+            // Tag types keep the flags compile-time, so __launch_bounds__ matches what is launched.
+            const auto launch = [&](auto small_k_tag, auto halve_iters_tag) {
+                constexpr bool c_small_k = decltype(small_k_tag)::value;
+                // Types the table does not promote would compile a second, identical kernel.
+                constexpr bool c_promoted =
+                    calc_nwarps(type, c_ncols_dst, MMVQ_PARAMETERS_GB10, false, true) !=
+                    calc_nwarps(type, c_ncols_dst, MMVQ_PARAMETERS_GB10, false, false);

-            if (use_small_k) {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
-                                                                        nsamples_dst, warp_size, table_id, true);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(
+                constexpr bool c_halve_iters = decltype(halve_iters_tag)::value && c_promoted;
+
+                const std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                              nsamples_dst, warp_size, table_id, c_small_k, c_halve_iters);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, c_small_k, c_halve_iters>(
                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
                    stream);
+            };
+
+            if (should_use_small_k(c_ncols_dst)) {
+                launch(std::true_type{},  std::false_type{});
+            } else if (should_halve_iters()) {
+                launch(std::false_type{}, std::true_type{});
            } else {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
-                                                                        nsamples_dst, warp_size, table_id);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
-                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
-                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
-                    stream);
+                launch(std::false_type{}, std::false_type{});
            }
        } break;
        case 2: {
@@ -54,8 +54,6 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const float alpha = 1.0f;
    const float beta = 0.0f;

-    CUBLAS_CHECK(cublasSetStream(handle, stream));
-
    const int64_t lda = nb01 / sizeof(float);
    const int64_t ldc = nb1  / sizeof(float);

@@ -53,6 +53,7 @@ static __global__ void rope_norm(const T *            x,
                                 const int            s2,
                                 const int            s3,
                                 const int            n_dims,
+                                 const int            n_offs,
                                 const int32_t *      pos,
                                 const float          freq_scale,
                                 const float          ext_factor,
@@ -61,7 +62,8 @@ static __global__ void rope_norm(const T *            x,
                                 const float          theta_scale,
                                 const float *        freq_factors,
                                 const int64_t *      row_indices,
-                                 const int            set_rows_stride) {
+                                 const int            set_rows_stride,
+                                 const bool           inplace) {
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

    if (i0 >= ne00) {
@@ -92,19 +94,24 @@ static __global__ void rope_norm(const T *            x,
            ggml_cuda_memcpy_1<4>(dst + idst, &v);
        }
    };
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        store_coaelsced(x[ix + 0], x[ix + 1]);
        return;
    }

-    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+    const int iw = i0 - n_offs; // relative idx

-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float theta_base = pos[i2]*powf(theta_scale, iw/2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw/2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor, cos_theta, sin_theta);

    const float x0 = x[ix + 0];
    const float x1 = x[ix + 1];
@@ -125,6 +132,7 @@ static __global__ void rope_neox(const T *            x,
                                 const int            s2,
                                 const int            s3,
                                 const int            n_dims,
+                                 const int            n_offs,
                                 const int32_t *      pos,
                                 const float          freq_scale,
                                 const float          ext_factor,
@@ -133,7 +141,8 @@ static __global__ void rope_neox(const T *            x,
                                 const float          theta_scale,
                                 const float *        freq_factors,
                                 const int64_t *      row_indices,
-                                 const int            set_rows_stride) {
+                                 const int            set_rows_stride,
+                                 const bool           inplace) {
    ggml_cuda_pdl_lc();
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

@@ -158,27 +167,33 @@ static __global__ void rope_neox(const T *            x,
        idst += row_indices[i2] * set_rows_stride;
    }

-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        dst[idst + i0 / 2 + 0] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 0]);
        dst[idst + i0 / 2 + 1] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 1]);

        return;
    }

-    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+    const int iw = i0 - n_offs; // relative idx

-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float theta_base = pos[i2]*powf(theta_scale, iw/2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw/2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor, cos_theta, sin_theta);

-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims/2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs/2 + 0];
+    const float x1 = x[ix + n_offs/2 + n_dims/2];

-    dst[idst + 0]          = ggml_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
-    dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
+    dst[idst + n_offs/2 + 0]          = ggml_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_offs/2 + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

 template <bool forward, bool has_ff, typename T>
@@ -194,6 +209,7 @@ static __global__ void rope_multi(const T *            x,
                                  const int            s2,
                                  const int            s3,
                                  const int            n_dims,
+                                  const int            n_offs,
                                  const int32_t *      pos,
                                  const float          freq_scale,
                                  const float          ext_factor,
@@ -202,7 +218,8 @@ static __global__ void rope_multi(const T *            x,
                                  const float          theta_scale,
                                  const float *        freq_factors,
                                  const mrope_sections sections,
-                                  const bool           is_imrope) {
+                                  const bool           is_imrope,
+                                  const bool           inplace) {
    const int i0 = 2 * (blockDim.y * blockIdx.y + threadIdx.y);

    if (i0 >= ne00) {
@@ -219,52 +236,58 @@ static __global__ void rope_multi(const T *            x,
    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

    ggml_cuda_pdl_sync();
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
        dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];

        return;
    }

+    const int iw = i0 - n_offs; // relative idx
+
    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;
+    const int sector = (iw / 2) % sect_dims;

    float theta_base = 0.0;
    if (is_imrope) {
        if (sector % 3 == 1 && sector < 3 * sections.v[1]) {         // h
-            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) {  // w
-            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) {  // t
-            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * powf(theta_scale, iw / 2.0f);
        } else {
-            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, iw / 2.0f);
        }
    } else {
        if (sector < sections.v[0]) {
-            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sections.v[0] && sector < sec_w) {
-            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, iw / 2.0f);
        }
    }

-    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+    const float freq_factor = has_ff ? freq_factors[iw/2] : 1.0f;

    float cos_theta;
    float sin_theta;

-    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, cos_theta, sin_theta);
+    rope_yarn<forward>(theta_base/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor, cos_theta, sin_theta);

-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims/2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs/2 + 0];
+    const float x1 = x[ix + n_offs/2 + n_dims/2];

-    dst[idst + 0]        = x0*cos_theta - x1*sin_theta;
-    dst[idst + n_dims/2] = x0*sin_theta + x1*cos_theta;
+    dst[idst + n_offs/2 + 0]        = x0*cos_theta - x1*sin_theta;
+    dst[idst + n_offs/2 + n_dims/2] = x0*sin_theta + x1*cos_theta;
 }

 template <bool forward, bool has_ff, typename T>
@@ -344,6 +367,7 @@ static void rope_norm_cuda(const T *            x,
                           const int            s2,
                           const int            s3,
                           const int            n_dims,
+                           const int            n_offs,
                           const int            nr,
                           const int32_t *      pos,
                           const float          freq_scale,
@@ -354,6 +378,7 @@ static void rope_norm_cuda(const T *            x,
                           const float *        freq_factors,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
+                           const bool           inplace,
                           cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -364,12 +389,12 @@ static void rope_norm_cuda(const T *            x,

    if (freq_factors == nullptr) {
        rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    } else {
        rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    }
 }

@@ -386,6 +411,7 @@ static void rope_neox_cuda(const T *            x,
                           const int            s2,
                           const int            s3,
                           const int            n_dims,
+                           const int            n_offs,
                           const int            nr,
                           const int32_t *      pos,
                           const float          freq_scale,
@@ -396,6 +422,7 @@ static void rope_neox_cuda(const T *            x,
                           const float *        freq_factors,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
+                           const bool           inplace,
                           cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -407,12 +434,12 @@ static void rope_neox_cuda(const T *            x,

    if (freq_factors == nullptr) {
        ggml_cuda_kernel_launch(rope_neox<forward, false, T, D>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    } else {
        ggml_cuda_kernel_launch(rope_neox<forward, true, T, D>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    }
 }

@@ -429,6 +456,7 @@ static void rope_multi_cuda(const T *            x,
                            const int            s2,
                            const int            s3,
                            const int            n_dims,
+                            const int            n_offs,
                            const int            nr,
                            const int32_t *      pos,
                            const float          freq_scale,
@@ -439,6 +467,7 @@ static void rope_multi_cuda(const T *            x,
                            const float *        freq_factors,
                            const mrope_sections sections,
                            const bool           is_imrope,
+                            const bool           inplace,
                            cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -450,13 +479,13 @@ static void rope_multi_cuda(const T *            x,
    if (freq_factors == nullptr) {
        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
        ggml_cuda_kernel_launch(rope_multi<forward, false, T>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope, inplace);
    } else {
        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
        ggml_cuda_kernel_launch(rope_multi<forward, true, T>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope, inplace);
    }
 }

@@ -552,8 +581,12 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    const int mode       = ((int32_t *) dst->op_params)[2];
    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];
    mrope_sections sections;

+    // when dst aliases src0, the channels outside the rotated window already hold the correct data
+    const bool inplace = dst_d == src0->data;
+
    // RoPE alteration for extended context
    float freq_base;
    float freq_scale;
@@ -581,6 +614,7 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,

    if (is_vision) {
        GGML_ASSERT(n_dims == ne00/2);
+        GGML_ASSERT(n_offs == 0); // offset not supported for vision, as the rotated pairs span the whole row
    }

    const int32_t * pos = (const int32_t *) src1_d;
@@ -597,31 +631,31 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    if (is_neox) {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                  set_rows_stride, stream);
+                                                  set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                 set_rows_stride, stream);
+                                                 set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                set_rows_stride, stream);
+                                                set_rows_stride, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
    } else if (is_mrope && !is_vision) {
        if (src0->type == GGML_TYPE_F32) {
            rope_multi_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
-                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, stream);
+                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16) {
            rope_multi_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
-                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, stream);
+                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -640,19 +674,19 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    } else {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                  set_rows_stride, stream);
+                                                  set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                 set_rows_stride, stream);
+                                                 set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
-                                                set_rows_stride, stream);
+                                                set_rows_stride, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -65,15 +65,13 @@ static void solve_tri_f32_cublas(ggml_backend_cuda_context & ctx,
    get_batch_pointers<<<(total_batches + 255) / 256, 256, 0, stream>>>(A, X, A_ptrs_dev, X_ptrs_dev, ne02,
                                                                        total_batches, s02, s03, s2, s3);

-    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
-
    // Yes, this is necessary, without this we get RMSE errors
-    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_DEFAULT_MATH));
-    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(id), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(), CUBLAS_DEFAULT_MATH));
+    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
                                    CUBLAS_DIAG_NON_UNIT, k, n, &alpha, A_ptrs_dev, n, X_ptrs_dev, k, total_batches));

    // revert to standard mode from common.cuh
-    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_TF32_TENSOR_OP_MATH));
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(), CUBLAS_TF32_TENSOR_OP_MATH));

    GGML_UNUSED_VARS(s12, s13);
 }
@@ -632,7 +632,6 @@ static void ssm_scan_ssd_f32_cuda(
    // Step 3: chunked SSD loop
    // Per chunk: pre_matmul (incl. M) + 4 cuBLAS (CB, Y, S@C, state update) + scale_state
    cublasHandle_t handle = ctx.cublas_handle();
-    CUBLAS_CHECK(cublasSetStream(handle, stream));
    const float alpha_one  = 1.0f;
    const float beta_zero  = 0.0f;
    const float beta_one   = 1.0f;
@@ -1061,9 +1061,11 @@ static bool ggml_backend_et_device_supports_op(ggml_backend_dev_t dev, const ggm
                const bool zero_view_offset = op->src[0]->view_src == nullptr || op->src[0]->view_offs == 0;
                const bool has_sections = ggml_get_op_params_i32(op, 11) > 0 || ggml_get_op_params_i32(op, 12) > 0 ||
                                          ggml_get_op_params_i32(op, 13) > 0;
+                // FIXME: support ggml_rope_set_offset
+                const bool zero_rot_offset  = ggml_get_op_params_i32(op, 15) == 0;

                supported =
-                    zero_view_offset && ndims <= 512 &&
+                    zero_view_offset && zero_rot_offset && ndims <= 512 &&
                    (is_normal || (is_neox && ndims % 16 == 0) || (is_imrope && ndims % 16 == 0 && has_sections));
            } else {
                supported = false;
@@ -3180,6 +3180,11 @@ static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * s
 static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
    const int32_t * op_params = &op->op_params[0];

+    // ggml_rope_set_offset: HVX kernels need a VLEN-aligned window start (32 f32 elems)
+    if (op_params[15] % 32 != 0) {
+        return false;
+    }
+
    int mode = op_params[2];

    // n_dims == ne0/2, so the rotation spans the full row
@@ -132,8 +132,8 @@ struct hmx_fa_context {
    __fp16 *     vtcm_v_tiles[2];      // V tiles (column-major, double-buffered)
    __fp16 *     vtcm_s_tiles[2];      // S = QK^T [g_br, Bc] (double-buffered)
    __fp16 *     vtcm_p_tiles[2];      // P = softmax(S) [g_br, Bc]
-    __fp16 *     vtcm_d_tiles;         // Diagonal rescale [g_br, g_br]
-    __fp16 *     vtcm_d_inv_l;         // Diagonal rescale (1/l) [g_br, g_br]
+    __fp16 *     vtcm_d_tiles[2];      // Diagonal rescale, g_br/32 packed diagonal tiles (double-buffered)
+    __fp16 *     vtcm_d_inv_l;         // Diagonal rescale (1/l), same packed layout
    HVX_Vector * vtcm_m_vec;           // Row max [g_br]
    HVX_Vector * vtcm_l_vec;           // Row sum [g_br]
    HVX_Vector * vtcm_s_rowmax;        // Softmax intermediate [g_br]
@@ -782,13 +782,14 @@ static void fa_q_load_thread(unsigned int n, unsigned int i, void * data) {
            }
        }

-        // Initialize vtcm_d_tiles and vtcm_d_inv_l to 0
+        // Zero the whole rescale region: vtcm_d_tiles[0], the optional vtcm_d_tiles[1]
+        // and vtcm_d_inv_l are equal-sized and allocated back to back, so one run covers
+        // them all.  The scatter only ever writes the diagonal, ignore the rest.
        const size_t d_bytes_per_t = hex_align_up(d_tile_bytes / n, 128);
        const size_t d_start       = i * d_bytes_per_t;
        const size_t d_end         = hex_smin(d_start + d_bytes_per_t, d_tile_bytes);
        if (d_start < d_tile_bytes) {
-            hvx_splat_u8_a((char *) factx->vtcm_d_tiles + d_start, 0, d_end - d_start);
-            hvx_splat_u8_a((char *) factx->vtcm_d_inv_l + d_start, 0, d_end - d_start);
+            hvx_splat_u8_a((char *) factx->vtcm_d_tiles[0] + d_start, 0, d_end - d_start);
        }
    }

@@ -1432,17 +1433,19 @@ static inline void fa_softmax_impl(
        const HVX_VectorPred q_32_mask = Q6_Q_vsetq_R(32 * sizeof(__fp16));
        HVX_Vector           v_exp_m_diff = exp_m_diff_f16;

+        __fp16 * const d_tiles_out = factx->vtcm_d_tiles[args->buf_idx];
+
        size_t t0 = r_vec_idx * 2;
        if (t0 < args->n_row_tiles) {
            const HVX_Vector v_content = v_exp_m_diff;
-            __fp16 *         out_base  = factx->vtcm_d_tiles + t0 * (args->n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+            __fp16 *         out_base  = d_tiles_out + t0 * HMX_FP16_TILE_N_ELMS;
            Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
        }

        size_t t1 = r_vec_idx * 2 + 1;
        if (t1 < args->n_row_tiles) {
            const HVX_Vector v_content = Q6_V_vror_VR(v_exp_m_diff, 64);
-            __fp16 *         out_base  = factx->vtcm_d_tiles + t1 * (args->n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+            __fp16 *         out_base  = d_tiles_out + t1 * HMX_FP16_TILE_N_ELMS;
            Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
        }
    }
@@ -1506,7 +1509,7 @@ static __attribute__((noinline)) void fa_build_d_diag_inv_l(struct hmx_fa_contex
            v_content = Q6_V_vror_VR(v_content, 64);
        }

-        __fp16 * out_base = factx->vtcm_d_inv_l + i * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        __fp16 * out_base = factx->vtcm_d_inv_l + i * HMX_FP16_TILE_N_ELMS;
        Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
    }
 }
@@ -1615,7 +1618,7 @@ static void hmx_fa_o_update_worker(void * data) {
    const size_t o_stride = n_row_tiles_g_br * HMX_FP16_TILE_N_ELMS;
    const size_t v_stride = n_tiles_per_bc * HMX_FP16_TILE_N_ELMS;
    for (size_t r = 0; r < n_row_tiles; ++r) {
-        const __fp16 * d_diag     = d_tiles + r * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        const __fp16 * d_diag     = d_tiles + r * HMX_FP16_TILE_N_ELMS;
        const __fp16 * p_tile_in  = p_tiles + (r * n_tiles_per_bc) * HMX_FP16_TILE_N_ELMS;
        const __fp16 * o_rc       = o_prev + r * HMX_FP16_TILE_N_ELMS;
        const __fp16 * v_tile_in  = v_tiles;
@@ -1654,7 +1657,7 @@ static void hmx_fa_o_norm_worker(void * data) {
    asm volatile(HMX_SET_BIAS("%0") :: "r"((unsigned int)job->hmx_scales));
    const size_t o_stride = n_row_tiles_g_br * HMX_FP16_TILE_N_ELMS;
    for (size_t r = 0; r < n_row_tiles; ++r) {
-        const __fp16 * d_diag = d_tiles + r * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        const __fp16 * d_diag = d_tiles + r * HMX_FP16_TILE_N_ELMS;
        const __fp16 * o_rc = o_prev + r * HMX_FP16_TILE_N_ELMS;
        __fp16 *       o_out = o_curr + r * DV_tiles * HMX_FP16_TILE_N_ELMS;

@@ -1882,7 +1885,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    factx.vtcm_s_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_s_tiles[1], pipeline);
    factx.vtcm_p_tiles[0]     = VTCM_LAYOUT_PTR(__fp16, base, L.off_p_tiles[0]);
    factx.vtcm_p_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_p_tiles[1], pipeline);
-    factx.vtcm_d_tiles        = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_tiles);
+    factx.vtcm_d_tiles[0]     = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_tiles[0]);
+    factx.vtcm_d_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_d_tiles[1], pipeline);
    factx.vtcm_d_inv_l        = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_inv_l);
    factx.vtcm_m_vec          = VTCM_LAYOUT_PTR(HVX_Vector, base, L.off_m_vec);
    factx.vtcm_l_vec          = VTCM_LAYOUT_PTR(HVX_Vector, base, L.off_l_vec);
@@ -2039,7 +2043,30 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            }
                        }

-                        // ---- 3. Pop and run K-prep for next block & push next QK-dot ----
+                        // ---- 3. Start HMX O update for block kv_blk - 1 (reads P[1 - buf_idx], V[1 - buf_idx], D) ----
+                        // O update relys on the previous block's P and V tiles.
+                        // O update MUST be pushed before the next block's QK-dot: hmx_queue_pop() retires the
+                        // oldest descriptor, so push order alone decides which pop waits for which job.
+                        // If OU went in after QK(i+1), the pop below would retire QK(i+1) and leave
+                        // OU(i-1) in flight into the next iteration, where V-prep overwrites V[prev_buf].
+                        if (kv_blk > 0) {
+                            const size_t prev_buf        = 1 - buf_idx;
+                            ou_job[prev_buf].o_curr      = o_tile_curr;
+                            ou_job[prev_buf].o_prev      = o_tile_prev;
+                            ou_job[prev_buf].p_tiles     = factx.vtcm_p_tiles[prev_buf];
+                            ou_job[prev_buf].v_tiles     = factx.vtcm_v_tiles[prev_buf];
+                            ou_job[prev_buf].d_tiles     = factx.vtcm_d_tiles[prev_buf];
+                            ou_job[prev_buf].hmx_scales  = factx.vtcm_hmx_scales_id;
+                            ou_job[prev_buf].n_row_tiles = n_row_tiles;
+                            ou_job[prev_buf].n_col_tiles =
+                                hmx_ceil_div(hex_smin(Bc, nek1 - (kv_blk - 1) * Bc), HMX_FP16_TILE_N_COLS);
+                            ou_job[prev_buf].n_row_tiles_g_br = n_row_tiles_g_br;
+                            ou_job[prev_buf].n_tiles_per_bc   = n_tiles_per_bc;
+                            ou_job[prev_buf].DV               = DV;
+                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[prev_buf]));
+                        }
+
+                        // ---- 4. Pop and run K-prep for next block & push next QK-dot ----
                        if (kv_blk + 1 < factx.n_kv_blocks) {
                            const uint32_t next_start = (kv_blk + 1) * Bc;
                            const uint32_t next_rows  = hex_smin(Bc, nek1 - next_start);
@@ -2059,10 +2086,10 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_qk_dot_worker, &qk_job[next_buf]));
                        }

-                        // ---- 4. Wait for current block's QK-dot to finish ----
+                        // ---- 5. Wait for current block's QK-dot to finish ----
                        hmx_queue_pop(hmx_q);

-                        // ---- 5. Phase 2: softmax + build_D ----
+                        // ---- 6. Phase 2: softmax + build_D ----
                        fa_softmax_args_t sargs;
                        memset(&sargs, 0, sizeof(sargs));
                        sargs.factx                = &factx;
@@ -2085,23 +2112,6 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                        sargs.mask_vtcm_row_stride = factx.mask_buf_row_stride;
                        sargs.slopes               = factx.vtcm_slopes;

-                        // Start HMX O update for block kv_blk - 1 (reads P[1 - buf_idx], V[1 - buf_idx])
-                        if (kv_blk > 0) {
-                            const size_t prev_buf = 1 - buf_idx;
-                            ou_job[prev_buf].o_curr           = o_tile_curr;
-                            ou_job[prev_buf].o_prev           = o_tile_prev;
-                            ou_job[prev_buf].p_tiles          = factx.vtcm_p_tiles[prev_buf];
-                            ou_job[prev_buf].v_tiles          = factx.vtcm_v_tiles[prev_buf];
-                            ou_job[prev_buf].d_tiles          = factx.vtcm_d_tiles;
-                            ou_job[prev_buf].hmx_scales       = factx.vtcm_hmx_scales_id;
-                            ou_job[prev_buf].n_row_tiles      = n_row_tiles;
-                            ou_job[prev_buf].n_col_tiles      = hmx_ceil_div(hex_smin(Bc, nek1 - (kv_blk - 1) * Bc), HMX_FP16_TILE_N_COLS);
-                            ou_job[prev_buf].n_row_tiles_g_br = n_row_tiles_g_br;
-                            ou_job[prev_buf].n_tiles_per_bc   = n_tiles_per_bc;
-                            ou_job[prev_buf].DV               = DV;
-                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[prev_buf]));
-                        }
-
                        // Run Softmax on HVX (blocking call)
                        fa_phase_softmax_and_build_d(&factx, &sargs, n_row_tiles, n_row_tiles_g_br);

@@ -2128,7 +2138,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                        ou_job[0].o_prev           = o_tile_prev;
                        ou_job[0].p_tiles          = factx.vtcm_p_tiles[1 - buf_idx];
                        ou_job[0].v_tiles          = factx.vtcm_v_tiles[1 - buf_idx];
-                        ou_job[0].d_tiles          = factx.vtcm_d_tiles;
+                        ou_job[0].d_tiles          = factx.vtcm_d_tiles[1 - buf_idx];
                        ou_job[0].hmx_scales       = factx.vtcm_hmx_scales_id;
                        ou_job[0].n_row_tiles      = n_row_tiles;
                        ou_job[0].n_col_tiles      = last_cols;
@@ -2232,7 +2242,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                            ou_job.o_prev           = o_tile_prev;
                            ou_job.p_tiles          = factx.vtcm_p_tiles[0];
                            ou_job.v_tiles          = factx.vtcm_v_tiles[0];
-                            ou_job.d_tiles          = factx.vtcm_d_tiles;
+                            ou_job.d_tiles          = factx.vtcm_d_tiles[0];
                            ou_job.hmx_scales       = factx.vtcm_hmx_scales_id;
                            ou_job.n_row_tiles      = n_row_tiles;
                            ou_job.n_col_tiles      = n_col_tiles;
@@ -109,7 +109,7 @@ struct hmx_fa_vtcm_layout {
    size_t off_v_tiles[2];
    size_t off_s_tiles[2];
    size_t off_p_tiles[2];
-    size_t off_d_tiles;
+    size_t off_d_tiles[2];
    size_t off_d_inv_l;
    size_t off_m_vec;
    size_t off_l_vec;
@@ -125,7 +125,7 @@ struct hmx_fa_vtcm_layout {
    size_t q_tile_bytes;
    size_t o_tile_bytes;
    size_t s_tile_bytes;       // S and P tiles (same size)
-    size_t d_tile_bytes;
+    size_t d_tile_bytes;       // d_tiles[0..1] + d_inv_l, allocated back to back
    size_t m_line_bytes;       // one mask row
    size_t m_buf_slot_bytes;   // one dma_cache slot = align_up(Br * m_line_bytes, 4096)
    size_t col_vec_bytes;
@@ -149,7 +149,12 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
    const size_t k_tile_size  = hex_align_up(Bc   * DK   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
    const size_t v_tile_size  = hex_align_up(Bc   * DV   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
    const size_t s_tile_size  = hex_align_up(g_br * Bc   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
-    const size_t d_tile_size  = hex_align_up(g_br * g_br * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
+
+    // The rescale matrices are diagonal: the HMX kernels only ever load the g_br/32
+    // tiles that sit on the diagonal, so store just those, packed back to back with
+    // a stride of one tile.  The old [g_br, g_br] square layout allocated g_br/32
+    // times more than it used, which is also why a second D buffer was unaffordable.
+    const size_t d_tile_size = (g_br / HMX_FP16_TILE_N_ROWS) * HTP_FA_HMX_TILE_SIZE;

    const size_t q_dma_size   = hex_align_up(g_br * DK * (is_q_fp32 ? sizeof(float) : sizeof(__fp16)), 128);
    const size_t k_dma_size   = hex_align_up(Bc * hex_round_up(DK * sizeof(__fp16), 128), 128);
@@ -167,7 +172,8 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
    VTCM_LAYOUT_ALLOC(off, off_q_tiles,       q_tile_size);
    VTCM_LAYOUT_ALLOC(off, off_o_tiles[0],    o_tile_size);
    VTCM_LAYOUT_ALLOC(off, off_o_tiles[1],    o_tile_size);
-    VTCM_LAYOUT_ALLOC(off, off_d_tiles,       d_tile_size);
+    VTCM_LAYOUT_ALLOC(off, off_d_tiles[0],    d_tile_size);
+    VTCM_LAYOUT_ALLOC_OPTIONAL(off, off_d_tiles[1], d_tile_size, pipeline);
    VTCM_LAYOUT_ALLOC(off, off_d_inv_l,       d_tile_size);

    // Group B & C share start offset (Group B tiles must be 2KB aligned)
@@ -213,7 +219,10 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
    L->o_tile_bytes        = o_tile_size;
    L->col_vec_bytes       = col_vec_size;
    L->s_tile_bytes        = s_tile_size;
-    L->d_tile_bytes        = d_tile_size;
+    // Measured from the actual offsets rather than assumed to be N * d_tile_size, so
+    // that inserting a region between them (or adding padding to VTCM_LAYOUT_ALLOC)
+    // cannot silently leave the tail of the run unzeroed.
+    L->d_tile_bytes        = (L->off_d_inv_l + d_tile_size) - L->off_d_tiles[0];
    L->m_line_bytes        = m_line_size;
    L->m_buf_slot_bytes    = m_buf_slot;
    L->row_buf_stride      = row_vec_size / 128;
@@ -53,6 +53,7 @@

 struct htp_rope_context {
    int32_t n_dims;
+    int32_t n_offs;
    int32_t mode;
    int32_t n_ctx_orig;
    int32_t sections[4];
@@ -405,32 +406,40 @@ static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict

 static void inline rope_basic_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
    #pragma unroll(4)
    for (uint32_t i = 0; i < nr; i++) {
        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
        float * s = (float *) (src + i * rctx->src0_row_size_aligned);

-        hvx_rope_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);

        // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
        }
    }
 }

 static void inline rope_neox_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
    #pragma unroll(4)
    for (uint32_t i = 0; i < nr; i++) {
        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
        float * s = (float *) (src + i * rctx->src0_row_size_aligned);

-        hvx_rope_neox_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_neox_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);

        // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
        }
    }
 }
@@ -673,6 +682,7 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
    rctx.n_dims     = ((const int32_t *) op_params)[1];
    rctx.mode       = ((const int32_t *) op_params)[2];
    rctx.n_ctx_orig = ((const int32_t *) op_params)[4];
+    rctx.n_offs     = ((const int32_t *) op_params)[15];

    memcpy(&rctx.freq_base,   (int32_t *) op_params + 5,  sizeof(float));
    memcpy(&rctx.freq_scale,  (int32_t *) op_params + 6,  sizeof(float));
@@ -24,62 +24,119 @@ if (GGML_METAL_NDEBUG)
 endif()

 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
+set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
+set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
+
+set(METALLIB_KERNEL_SOURCES
+    kernels/fa.metal
+    kernels/mul_mv.metal
+    kernels/mul_mm.metal
+    kernels/quantize.metal
+    kernels/softmax.metal
+    kernels/norm.metal
+    kernels/unary.metal
+    kernels/binbcast.metal
+    kernels/reduce.metal
+    kernels/tri.metal
+    kernels/ssm.metal
+    kernels/wkv.metal
+    kernels/gated_delta_net.metal
+    kernels/solve_tri.metal
+    kernels/rope.metal
+    kernels/conv.metal
+    kernels/upscale.metal
+    kernels/argsort.metal
+    kernels/pool.metal
+    kernels/misc.metal
+)
+
 if (GGML_METAL_EMBED_LIBRARY)
    enable_language(ASM)

    add_compile_definitions(GGML_METAL_EMBED_LIBRARY)

-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")

    file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")

-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(kind ${src} NAME_WE)
+        # symbol names must be valid C identifiers ('-' is not allowed)
+        string(REPLACE "-" "_" kind_sym ${kind})

-    add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
-    )
+        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
+        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
+        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")

-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # only prepend headers that this source actually includes
+        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
+        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
+        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
+        if(_has_dequantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
+        endif()
+        if(_has_quantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
+        endif()
+
+        add_custom_command(
+            OUTPUT "${ASM}"
+            # Step 1: concatenate shared headers + this kernel source
+            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
+            # Step 2: remove internal #include and #pragma once
+            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
+            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
+            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
+            # Step 4: inline ggml-metal-impl.h
+            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
+            # Step 5: emit an asm chunk with kind-specific start/end symbols
+            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
+            #   section name is limited to 16 chars so we keep it shared
+            #   across kinds (__ggml_metallib) and only vary the global symbols.
+            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
+            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
+            DEPENDS ../ggml-common.h ggml-metal-impl.h
+                    kernels/common.h kernels/dequantize.h kernels/quantize.h
+                    kernels/${kind}.metal
+            COMMENT "Generate embedded Metal library for ${kind}"
+            VERBATIM
+        )
+
+        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
+    endforeach()
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
 else()
-    # copy metal files to bin directory
+    # copy header files to bin directory
    configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
    configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)

+    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
+    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
+    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
+    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
+
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
+    endforeach()
+
    if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: disabling fast math is needed in order to pass tests/test-backend-ops
        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
        #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
        # note: adding -g causes segmentation fault during compile
-        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
        set(XC_FLAGS -fno-fast-math -fno-inline)
    else()
        set(XC_FLAGS -O3)
    endif()

-    # Append macOS metal versioning flags
    if (GGML_METAL_MACOSX_VERSION_MIN)
        message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
        list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -90,35 +147,46 @@ else()
        list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
    endif()

+    # Compile each kernel source to .air, then link into default.metallib
+    set(AIR_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(name ${src} NAME_WE)
+        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
+        list(APPEND AIR_FILES ${AIR})
+        add_custom_command(
+            OUTPUT ${AIR}
+            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling ${src}"
+            VERBATIM
+        )
+    endforeach()
+
    add_custom_command(
        OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
+        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
+        DEPENDS ${AIR_FILES}
+        COMMENT "Linking Metal kernels into default.metallib"
+    )

-    # FIXME: only add to the ggml-metal target?
    add_custom_target(
        ggml-metal-lib ALL
        DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
+    )
 endif() # GGML_METAL_EMBED_LIBRARY

 if (NOT GGML_METAL_EMBED_LIBRARY)
    install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )

-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
+    install(
+        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        DESTINATION ${CMAKE_INSTALL_BINDIR}
+    )
 endif()
@@ -1409,6 +1409,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_p
    return res;
 }

+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    char base[256];
+
+    snprintf(base, 256, "kernel_flash_attn_ext_kv_%s_f16", ggml_type_name(op->src[1]->type));
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, base);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, base, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_blk(
        ggml_metal_library_t lib,
        const struct ggml_tensor * op,
@@ -1460,7 +1477,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
-        int32_t nsg) {
+        int32_t nsg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20) {
    assert(op->op == GGML_OP_FLASH_ATTN_EXT);

    char base[256];
@@ -1469,15 +1489,14 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
    const int32_t dk = (int32_t) op->src[1]->ne[0];
    const int32_t dv = (int32_t) op->src[2]->ne[0];

-    const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
-    const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
+    const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);

    // do bounds checks for the mask?
    const bool bc_mask = op->src[3] && (op->src[3]->ne[1] % 8 != 0);

    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
            "flash_attn_ext",
-            ggml_type_name(op->src[1]->type),
+            type,
            dk,
            dv);

@@ -1526,7 +1545,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
        bool    has_scap,
        bool    has_kvpad,
        int32_t nsg,
-        int32_t nwg) {
+        int32_t nwg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20) {
    assert(op->op == GGML_OP_FLASH_ATTN_EXT);

    char base[256];
@@ -1535,12 +1557,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
    const int32_t dk = (int32_t) op->src[1]->ne[0];
    const int32_t dv = (int32_t) op->src[2]->ne[0];

-    const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
-    const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
+    const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);

    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
            "flash_attn_ext_vec",
-            ggml_type_name(op->src[1]->type),
+            type,
            dk,
            dv);

@@ -176,6 +176,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_mask,
        int32_t ncpsg);

+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op);
+
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_blk(
        ggml_metal_library_t lib,
        const struct ggml_tensor * op,
@@ -190,7 +194,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
-        int32_t nsg);
+        int32_t nsg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20);

 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec(
        ggml_metal_library_t lib,
@@ -201,7 +208,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_scap,
        bool    has_kvpad,
        int32_t nsg,
-        int32_t nwg);
+        int32_t nwg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20);

 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(
        ggml_metal_library_t lib,
@@ -95,8 +95,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
    return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }

+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;

    ggml_metal_device_t dev;
    ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -104,160 +159,376 @@ struct ggml_metal_library {
    NSLock * lock;
 };

-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}

-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];

-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }

-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }

-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}

-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];

-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }

-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }

-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
            }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
        }

-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }

-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}

-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];

-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];

-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
            }

-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;

-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
            @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                MTLCompileOptions * options = [MTLCompileOptions new];
                options.preprocessorMacros = prep;

-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];

-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
            }
        }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
    }

-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);

-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
    res->dev       = dev;
    res->pipelines = ggml_metal_pipelines_init();
    res->lock      = [NSLock new];

+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
    return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }

 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -319,10 +590,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
        return NULL;
    }

-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];

    return res;
 }
@@ -332,8 +604,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
        return;
    }

-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
    }

    ggml_metal_pipelines_free(lib->pipelines);
@@ -394,11 +672,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_

        GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);

+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
        id<MTLFunction> mtl_function;
        if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
        } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
        }
        if (!mtl_function) {
            [lib->lock unlock];
@@ -329,6 +329,7 @@ typedef struct {
    uint64_t nb3;
    int32_t  n_past;
    int32_t  n_dims;
+    int32_t  n_offs;
    int32_t  n_ctx_orig;
    float    freq_base;
    float    freq_scale;
@@ -341,8 +342,21 @@ typedef struct {
    int32_t  sect_2;
    int32_t  sect_3;
    bool     src2;
+    bool     inplace;
 } ggml_metal_kargs_rope;

+typedef struct {
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  nblocks;
+} ggml_metal_kargs_flash_attn_ext_kv_f16;
+
 typedef struct {
    int32_t  ne11;
    int32_t  ne_12_2; // assume K and V are same shape
@@ -2801,6 +2801,51 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const ggml_tensor * op) {
    return (ne01 < 20) && (ne00 % 32 == 0);
 }

+// ref: https://github.com/ggml-org/llama.cpp/pull/27390
+// dequantize the quantized KV cache to F16 before running the F16 flash attention kernels
+static bool ggml_metal_op_flash_attn_ext_use_kv_f16(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    // depending on compute/bandwidth ratio, dequant to f16 kv is not always beneficial
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27390#issuecomment-5355152767
+    // TODO: tune per device
+    if (op->src[0]->ne[1] < 32) {
+        return false;
+    }
+
+    switch (op->src[1]->type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+// in some models (e.g. MLA-based), V is a view of K (the first ne20 elements of each K row);
+// the dequantized V is then a view of the dequantized K and does not need its own dequant or scratch
+// - ref: https://github.com/ggml-org/llama.cpp/pull/13435
+static bool ggml_metal_op_flash_attn_ext_v_is_view_of_k(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    const ggml_tensor * K = op->src[1];
+    const ggml_tensor * V = op->src[2];
+
+    return V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs));
+}
+
+// size of the F16 dequantized K tensor; the dequantized V tensor follows it in the same scratch buffer
+static size_t ggml_metal_op_flash_attn_ext_kv_f16_k_size(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+
+    return GGML_PAD(sizeof(ggml_fp16_t)*(size_t) ne10*ne11*ne12*ne13, 16);
+}
+
 size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
    assert(op->op == GGML_OP_FLASH_ATTN_EXT);

@@ -2816,6 +2861,18 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
    size_t res = 0;

    const bool has_mask = op->src[3] != nullptr;
+    const bool use_kv_f16 = ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+
+    // when the KV is dequantized to F16, the pad kernel copies the tail chunk from the F16 scratch buffer
+    // note: when V is a view of K, the dequantized V is read from the dequantized K with K's row stride
+    const bool v_is_view_of_k = use_kv_f16 && ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+    uint64_t nb11_pad = nb11;
+    uint64_t nb21_pad = nb21;
+
+    if (use_kv_f16) {
+        nb11_pad = sizeof(ggml_fp16_t)*ne10;
+        nb21_pad = sizeof(ggml_fp16_t)*(v_is_view_of_k ? ne10 : ne20);
+    }

    // note: the non-vec kernel requires more extra memory, so always reserve for it
    GGML_ASSERT(OP_FLASH_ATTN_EXT_NCPSG >= OP_FLASH_ATTN_EXT_VEC_NCPSG);
@@ -2828,8 +2885,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {

        if (has_kvpad) {
            res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
-                nb11*ne12*ne13 +
-                nb21*ne22*ne23 +
+                nb11_pad*ne12*ne13 +
+                nb21_pad*ne22*ne23 +
                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
        }
    } else {
@@ -2838,8 +2895,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {

        if (has_kvpad) {
            res += OP_FLASH_ATTN_EXT_NCPSG*(
-                nb11*ne12*ne13 +
-                nb21*ne22*ne23 +
+                nb11_pad*ne12*ne13 +
+                nb21_pad*ne22*ne23 +
                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
        }
    }
@@ -2915,6 +2972,29 @@ size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
    return res;
 }

+size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    // note: always reserve the temp buffer to avoid graph reallocations
+    //if (!ggml_metal_op_flash_attn_ext_use_kv_f16(op)) {
+    //    return 0;
+    //}
+
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+
+    const size_t k_size = ggml_metal_op_flash_attn_ext_kv_f16_k_size(op);
+
+    // when V is a view of K, the dequantized V is a view of the dequantized K
+    const bool v_is_view_of_k = ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+    if (v_is_view_of_k) {
+        return k_size;
+    }
+
+    const size_t v_size = GGML_PAD(sizeof(ggml_fp16_t)*(size_t) ne20*ne21*ne22*ne23, 16);
+
+    return k_size + v_size;
+}
+
 int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -2989,6 +3069,111 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
    ggml_metal_buffer_id bid_tmp = bid_blk;
    bid_tmp.offs += ggml_metal_op_flash_attn_ext_extra_blk(op);

+    ggml_metal_buffer_id bid_kv_f16 = bid_tmp;
+    bid_kv_f16.offs += ggml_metal_op_flash_attn_ext_extra_tmp(op);
+
+    const bool use_kv_f16 = ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+
+    ggml_metal_buffer_id bid_k = bid_src1;
+    ggml_metal_buffer_id bid_v = bid_src2;
+
+    uint64_t nb10_attn = nb10;
+    uint64_t nb11_attn = nb11;
+    uint64_t nb12_attn = nb12;
+    uint64_t nb13_attn = nb13;
+    uint64_t nb20_attn = nb20;
+    uint64_t nb21_attn = nb21;
+    uint64_t nb22_attn = nb22;
+    uint64_t nb23_attn = nb23;
+
+    if (use_kv_f16) {
+        assert(ggml_metal_op_flash_attn_ext_extra_kv_f16(op) != 0);
+
+        const bool v_is_view_of_k = ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+
+        const int64_t nblocks1_64 = (ne10/ggml_blck_size(op->src[1]->type))*(int64_t) ne11*ne12*ne13;
+        GGML_ASSERT(nblocks1_64 <= INT32_MAX);
+        const int32_t nblocks1 = nblocks1_64;
+
+        ggml_metal_buffer_id bid_v_f16 = bid_kv_f16;
+        bid_v_f16.offs += ggml_metal_op_flash_attn_ext_kv_f16_k_size(op);
+
+        auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(lib, op);
+        const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline0), 256);
+
+        // K
+        ggml_metal_kargs_flash_attn_ext_kv_f16 args_k = {
+            /*.ne0    =*/ ne10,
+            /*.ne1    =*/ ne11,
+            /*.ne2    =*/ ne12,
+            /*.ne3    =*/ ne13,
+            /*.nb0    =*/ nb10,
+            /*.nb1    =*/ nb11,
+            /*.nb2    =*/ nb12,
+            /*.nb3    =*/ nb13,
+            /*.nblocks =*/ nblocks1,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline0);
+        ggml_metal_encoder_set_bytes   (enc, &args_k, sizeof(args_k), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1,        1);
+        ggml_metal_encoder_set_buffer  (enc, bid_kv_f16, 2);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, (nblocks1 + nth - 1)/nth, 1, 1, nth, 1, 1);
+
+        // V (skip when V is a view of K: the dequantized V is a view of the dequantized K)
+        if (!v_is_view_of_k) {
+            const int64_t nblocks2_64 = (ne20/ggml_blck_size(op->src[2]->type))*(int64_t) ne21*ne22*ne23;
+            GGML_ASSERT(nblocks2_64 <= INT32_MAX);
+            const int32_t nblocks2 = nblocks2_64;
+
+            ggml_metal_kargs_flash_attn_ext_kv_f16 args_v = {
+                /*.ne0    =*/ ne20,
+                /*.ne1    =*/ ne21,
+                /*.ne2    =*/ ne22,
+                /*.ne3    =*/ ne23,
+                /*.nb0    =*/ nb20,
+                /*.nb1    =*/ nb21,
+                /*.nb2    =*/ nb22,
+                /*.nb3    =*/ nb23,
+                /*.nblocks =*/ nblocks2,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args_v, sizeof(args_v), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2,        1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v_f16,       2);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, (nblocks2 + nth - 1)/nth, 1, 1, nth, 1, 1);
+        }
+
+        // the pad and attention kernels read the dequantized KV
+        ggml_metal_op_concurrency_reset(ctx);
+
+        bid_k = bid_kv_f16;
+        bid_v = v_is_view_of_k ? bid_k : bid_v_f16;
+
+        // contiguous F16 layout of the dequantized K
+        nb10_attn = sizeof(ggml_fp16_t);
+        nb11_attn = nb10_attn*ne10;
+        nb12_attn = nb11_attn*ne11;
+        nb13_attn = nb12_attn*ne12;
+
+        // if V is a view of K, the dequantized V is read from the dequantized K with K's strides
+        if (v_is_view_of_k) {
+            nb20_attn = nb10_attn;
+            nb21_attn = nb11_attn;
+            nb22_attn = nb12_attn;
+            nb23_attn = nb13_attn;
+        } else {
+            // contiguous F16 layout of the dequantized V
+            nb20_attn = sizeof(ggml_fp16_t);
+            nb21_attn = nb20_attn*ne20;
+            nb22_attn = nb21_attn*ne21;
+            nb23_attn = nb22_attn*ne22;
+        }
+    }
+
    if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
        // half8x8 kernel
        const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
@@ -3009,12 +3194,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
                /*.ne11    =*/ne11,
                /*.ne_12_2 =*/ne12,
                /*.ne_12_3 =*/ne13,
-                /*.nb11    =*/nb11,
-                /*.nb12    =*/nb12,
-                /*.nb13    =*/nb13,
-                /*.nb21    =*/nb21,
-                /*.nb22    =*/nb22,
-                /*.nb23    =*/nb23,
+                /*.nb11    =*/nb11_attn,
+                /*.nb12    =*/nb12_attn,
+                /*.nb13    =*/nb13_attn,
+                /*.nb21    =*/nb21_attn,
+                /*.nb22    =*/nb22_attn,
+                /*.nb23    =*/nb23_attn,
                /*.ne31    =*/ne31,
                /*.ne32    =*/ne32,
                /*.ne33    =*/ne33,
@@ -3027,8 +3212,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

            ggml_metal_encoder_set_pipeline(enc, pipeline0);
            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
-            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
-            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_k,    1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v,    2);
            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);

@@ -3073,7 +3258,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            ggml_metal_op_concurrency_reset(ctx);
        }

-        const int is_q = ggml_is_quantized(op->src[1]->type) ? 1 : 0;
+        const int is_q = !use_kv_f16 && ggml_is_quantized(op->src[1]->type) ? 1 : 0;

        // 2*(2*ncpsg)
        // ncpsg soft_max values + ncpsg mask values
@@ -3104,6 +3289,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

        const size_t smem = FATTN_SMEM(nsg);

+        const int32_t ns10 = nb11_attn/nb10_attn;
+        const int32_t ns20 = nb21_attn/nb20_attn;
+
        ggml_metal_kargs_flash_attn_ext args = {
            /*.ne01          =*/ ne01,
            /*.ne02          =*/ ne02,
@@ -3114,14 +3302,14 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.ne11          =*/ ne11,
            /*.ne_12_2       =*/ ne12,
            /*.ne_12_3       =*/ ne13,
-            /*.ns10          =*/ int32_t(nb11/nb10),
-            /*.nb11          =*/ nb11,
-            /*.nb12          =*/ nb12,
-            /*.nb13          =*/ nb13,
-            /*.ns20          =*/ int32_t(nb21/nb20),
-            /*.nb21          =*/ nb21,
-            /*.nb22          =*/ nb22,
-            /*.nb23          =*/ nb23,
+            /*.ns10          =*/ ns10,
+            /*.nb11          =*/ nb11_attn,
+            /*.nb12          =*/ nb12_attn,
+            /*.nb13          =*/ nb13_attn,
+            /*.ns20          =*/ ns20,
+            /*.nb21          =*/ nb21_attn,
+            /*.nb22          =*/ nb22_attn,
+            /*.nb23          =*/ nb23_attn,
            /*.ne31          =*/ ne31,
            /*.ne32          =*/ ne32,
            /*.ne33          =*/ ne33,
@@ -3139,13 +3327,13 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.logit_softcap =*/ logit_softcap,
        };

-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, use_kv_f16, ns10, ns20);

        ggml_metal_encoder_set_pipeline(enc, pipeline);
        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
-        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
-        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_k,    2);
+        ggml_metal_encoder_set_buffer  (enc, bid_v,    3);
        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
        ggml_metal_encoder_set_buffer  (enc, bid_pad,  6);
@@ -3177,12 +3365,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
                /*.ne11    =*/ne11,
                /*.ne_12_2 =*/ne12,
                /*.ne_12_3 =*/ne13,
-                /*.nb11    =*/nb11,
-                /*.nb12    =*/nb12,
-                /*.nb13    =*/nb13,
-                /*.nb21    =*/nb21,
-                /*.nb22    =*/nb22,
-                /*.nb23    =*/nb23,
+                /*.nb11    =*/nb11_attn,
+                /*.nb12    =*/nb12_attn,
+                /*.nb13    =*/nb13_attn,
+                /*.nb21    =*/nb21_attn,
+                /*.nb22    =*/nb22_attn,
+                /*.nb23    =*/nb23_attn,
                /*.ne31    =*/ne31,
                /*.ne32    =*/ne32,
                /*.ne33    =*/ne33,
@@ -3195,8 +3383,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

            ggml_metal_encoder_set_pipeline(enc, pipeline0);
            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
-            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
-            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_k,    1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v,    2);
            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);

@@ -3242,6 +3430,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            }
        }

+        const int32_t ns10 = nb11_attn/nb10_attn;
+        const int32_t ns20 = nb21_attn/nb20_attn;
+
        ggml_metal_kargs_flash_attn_ext_vec args = {
            /*.ne01          =*/ ne01,
            /*.ne02          =*/ ne02,
@@ -3252,14 +3443,14 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.ne11          =*/ ne11,
            /*.ne_12_2       =*/ ne12,
            /*.ne_12_3       =*/ ne13,
-            /*.ns10          =*/ int32_t(nb11/nb10),
-            /*.nb11          =*/ nb11,
-            /*.nb12          =*/ nb12,
-            /*.nb13          =*/ nb13,
-            /*.ns20          =*/ int32_t(nb21/nb20),
-            /*.nb21          =*/ nb21,
-            /*.nb22          =*/ nb22,
-            /*.nb23          =*/ nb23,
+            /*.ns10          =*/ ns10,
+            /*.nb11          =*/ nb11_attn,
+            /*.nb12          =*/ nb12_attn,
+            /*.nb13          =*/ nb13_attn,
+            /*.ns20          =*/ ns20,
+            /*.nb21          =*/ nb21_attn,
+            /*.nb22          =*/ nb22_attn,
+            /*.nb23          =*/ nb23_attn,
            /*.ne31          =*/ ne31,
            /*.ne32          =*/ ne32,
            /*.ne33          =*/ ne33,
@@ -3277,15 +3468,15 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.logit_softcap =*/ logit_softcap,
        };

-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg, use_kv_f16, ns10, ns20);

        GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));

        ggml_metal_encoder_set_pipeline(enc, pipeline);
        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
-        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
-        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_k,    2);
+        ggml_metal_encoder_set_buffer  (enc, bid_v,    3);
        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);

@@ -3884,6 +4075,11 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
    const int sect_2 = ((const int32_t *) op->op_params)[13];
    const int sect_3 = ((const int32_t *) op->op_params)[14];

+    const int n_offs = ((const int32_t *) op->op_params)[15];
+
+    // when dst aliases src0, the channels outside the rotated window already hold the correct data
+    const bool inplace = op->data == op->src[0]->data;
+
    ggml_metal_kargs_rope args = {
        /*.ne00        =*/ ne00,
        /*.ne01        =*/ ne01,
@@ -3903,6 +4099,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
        /*.nb3         =*/ nb3,
        /*.n_past      =*/ n_past,
        /*.n_dims      =*/ n_dims,
+        /*.n_offs      =*/ n_offs,
        /*.n_ctx_orig  =*/ n_ctx_orig,
        /*.freq_base   =*/ freq_base,
        /*.freq_scale  =*/ freq_scale,
@@ -3915,6 +4112,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
        /* sect_2      =*/ sect_2,
        /* sect_3      =*/ sect_3,
        /* src2        =*/ op->src[2] != nullptr,
+        /* inplace     =*/ inplace,
    };

    auto pipeline = ggml_metal_library_get_pipeline_rope(lib, op);
@@ -42,6 +42,7 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_pad(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_blk(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
+size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const struct ggml_tensor * op);

 int ggml_metal_op_concat            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_repeat            (ggml_metal_op_t ctx, int idx);
@@ -225,6 +225,7 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
                res += ggml_metal_op_flash_attn_ext_extra_pad(tensor);
                res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
                res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_kv_f16(tensor);
            } break;
        case GGML_OP_CUMSUM:
        case GGML_OP_ARGSORT:
@@ -0,0 +1,232 @@
+#include "common.h"
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    // bitonic sort
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
+
+    // initialize indices
+    shmem_i32[col] = i00 + col;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ntg.x; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                } else {
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+
+    const int64_t i0 = ib*args.top_k;
+
+    // copy the result to dst without the padding
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
+    }
+}
+
+template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
+
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;
@@ -0,0 +1,228 @@
+#include "common.h"
+
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+constant bool  FC_bin_cb [[function_constant(FC_BIN + 3)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
+        constant ggml_metal_kargs_bin & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
+#define FC_CB FC_bin_cb
+
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
+        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
+
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
+
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
+
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
+
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
+
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
+
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            dst_row[i0] = res;
+        }
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
+        }
+    }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
+#undef FC_CB
+}
+
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
+
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half,   half,   half>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half4,  half4,  half4>;
+
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
+
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_repeat_bf16")]] kernel kernel_repeat_t kernel_repeat<bfloat>;
+#endif
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;
@@ -0,0 +1,126 @@
+#pragma once
+
+#include "ggml-metal-impl.h"
+
+#include <metal_stdlib>
+
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define PAD2(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#define FOR_UNROLL(x) _Pragma("clang loop unroll(full)") for (x)
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+//
+// cmd:
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/kernels/<src>.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/kernels/<src>.metal
+//
+#if __METAL_VERSION__ < 310 && defined(GGML_METAL_HAS_BF16)
+#undef GGML_METAL_HAS_BF16
+#endif
+
+#if defined(GGML_METAL_HAS_BF16)
+typedef matrix<bfloat, 4, 4> bfloat4x4;
+typedef matrix<bfloat, 2, 4> bfloat2x4;
+#endif
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline int best_index_int8(int n, constant float * 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 inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
+static inline float dot(float x, float y) {
+    return x*y;
+}
+
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+inline T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}
@@ -0,0 +1,723 @@
+#include "common.h"
+
+typedef void (im2col_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+          int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0;
+    const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1;
+
+    int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        while (in < args.N) {
+            pdst[offset_dst] = 0.0f;
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+
+            in += ntg[0];
+        }
+    } else {
+        int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw;
+
+        while (in < args.N) {
+            pdst[offset_dst] = x[offset_src];
+
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+            offset_src += ntg[0]*args.ofs0;
+
+            in += ntg[0];
+        }
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+typedef void (conv_transpose_1d_t)(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    // For output position j on the time axis, only input positions
+    //   i such that i*s0 <= j < i*s0 + K
+    // contribute -- i.e. i in [ceil((j - K + 1)/s0), floor(j/s0)]
+    // intersected with [0, IL-1]. That's at most ceil(K/s0) values
+    // (typically 2 for stride==K/2 transposed convs).
+    const int32_t j  = tgpig[0];
+    const int32_t s0 = args.s0;
+    const int32_t K  = args.K;
+    const int32_t IL = args.IL;
+
+    int32_t i_min;
+    {
+        int32_t a = j - K + 1;
+        i_min = a <= 0 ? 0 : (a + s0 - 1) / s0; // ceil(a/s0) for a>0
+    }
+    int32_t i_max = j / s0;
+    if (i_max > IL - 1) i_max = IL - 1;
+
+    float v = 0.0f;
+    if (i_min <= i_max) {
+        for (int64_t c = 0; c < args.IC; c++) {
+            const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+            const int32_t input_offset  = c * IL;
+
+            for (int32_t i = i_min; i <= i_max; i++) {
+                v += float(src0[kernel_offset + j - i * s0]) * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+
+typedef void (conv_transpose_2d_t)(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_2d(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const T * src0,
+        device const float * src1,
+        device        char * dst,
+        threadgroup float * shared_sum [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int64_t out_x = tgpig[0];
+    const int64_t out_y = tgpig[1];
+    const int64_t out_c = tgpig[2];
+
+    const int64_t kw = tpitg[0];
+    const int64_t kh = tpitg[1];
+
+    float v = 0.0f;
+
+    for (int64_t in_c = 0; in_c < args.IC; in_c++) {
+        int64_t in_y = out_y - kh;
+
+        if (in_y < 0 || in_y % args.s0) continue;
+
+        in_y /= args.s0;
+
+        if (in_y >= args.IH) continue;
+
+        int64_t in_x = out_x - kw;
+
+        if (in_x < 0 || in_x % args.s0) continue;
+
+        in_x /= args.s0;
+
+        if (in_x >= args.IW) continue;
+
+        const int64_t input_idx = (args.IW * args.IH) * in_c + (args.IW) * in_y + in_x;
+        const int64_t kernel_idx = (args.KH * args.KW * args.OC) * in_c + (args.KH * args.KW) * out_c + (args.KW) * kh + kw;
+
+        v += (float)src0[kernel_idx] * src1[input_idx];
+    }
+
+    const uint tid = tpitg.y * ntg.x + tpitg.x;
+    shared_sum[tid] = v;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tid == 0) {
+        float total = 0.0f;
+        const uint num_threads = ntg.x * ntg.y;
+        for (uint i = 0; i < num_threads; i++) {
+            total += shared_sum[i];
+        }
+
+        device float * dst_ptr = (device float *) (dst + out_x*args.nb0 + out_y * args.nb1 + out_c*args.nb2);
+        dst_ptr[0] = total;
+    }
+}
+
+template [[host_name("kernel_conv_transpose_2d_f32_f32")]]
+kernel void kernel_conv_transpose_2d<float>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_transpose_2d_f16_f32")]]
+kernel void kernel_conv_transpose_2d<half>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+// grid: x = C tile, y = OH, z = OW * N (for channel-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw_tiled(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t c = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (c >= args.C) {
+        return;
+    }
+
+    const int32_t oh = tgpig.y;
+    const int32_t own = tgpig.z;
+    const int32_t ow = own % args.OW;
+    const int32_t n  = own / args.OW;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    float acc = 0.0f;
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        const uint64_t w_base   = (uint64_t) c * args.nb02;
+        const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+// grid: x = OW tile, y = OH, z = C * N (for spatially-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t oh = tgpig.y;
+    const int32_t cn = tgpig.z;
+    const int32_t c  = cn % args.C;
+    const int32_t n  = cn / args.C;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const uint64_t w_base   = (uint64_t) c * args.nb02;
+    const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+    const int32_t ow = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (ow >= args.OW) {
+        return;
+    }
+
+    float acc = 0.0f;
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+template [[host_name("kernel_conv_2d_dw_f32_f32")]]
+kernel void kernel_conv_2d_dw<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_f16_f32")]]
+kernel void kernel_conv_2d_dw<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f32_f32")]]
+kernel void kernel_conv_2d_dw_tiled<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f16_f32")]]
+kernel void kernel_conv_2d_dw_tiled<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
@@ -0,0 +1,735 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    const int byte_offset = il * 4;  // il*16 elements = il*4 bytes (4 elements per byte)
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; i++) {
+        const uint8_t b = qs[byte_offset + i];
+        reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+        reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+        reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+        reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const uint8_t b = xb->qs[il];
+
+    float4 reg_f;
+    reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+    reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+    reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+    reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; ++i) {
+        reg_f[i] = float4(qs[4*il + i]) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
+    const float d = xb->d;
+
+    reg = (type4) (float4(qs[il]) * d);
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4x4>
+void dequantize_tq2_0(device const block_tq2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    // 2 bits per element, 4 elements per byte, 128 elements per 32-byte group
+    const short base = il * 16;
+    for (int k = 0; k < 16; k++) {
+        const int i = base + k;
+        const int byte = ((i >> 7) & 1) * 32 + (i & 31);
+        const int l = (i >> 5) & 3;
+        reg_f[k/4][k%4] = d * (float)(((qs[byte] >> (2*l)) & 3) - 1);
+    }
+
+    reg = (type4x4) reg_f;
+}
@@ -0,0 +1,250 @@
+#include "common.h"
+
+constant short FC_gated_delta_net_ne20 [[function_constant(FC_GATED_DELTA_NET + 0)]];
+constant short FC_gated_delta_net_ne30 [[function_constant(FC_GATED_DELTA_NET + 1)]];
+constant short FC_gated_delta_net_K    [[function_constant(FC_GATED_DELTA_NET + 2)]];
+
+#if 1
+template<short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+#define K   FC_gated_delta_net_K
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B (n_seqs)
+    const uint i21 = tgpig.y; // H (head)
+    const uint i20 = tgpig.x*NSG + ty; // row within S_v
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    // input state layout [S_v, S_v, H, n_seqs] (s0 only): per-seq stride is H*D.
+    // state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
+    const uint state_in_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+    device const float * s_ptr = (device const float *) (s) + state_in_base;
+
+    float ls[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        ls[j] = s_ptr[is];
+    }
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
+
+    // output state base offset: after attention scores
+    const uint attn_size = args.ne22 * args.ne21 * S_v * args.ne23;
+    // output state per-slot size: S_v * S_v * H * n_seqs
+    const uint state_size_per_snap = S_v * S_v * args.ne21 * args.ne23;
+    // per-(seq,head) offset within a slot
+    const uint state_out_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+
+    for (short t = 0; t < args.ne22; t++) {
+        float s_k = 0.0f;
+
+        if (G == 1) {
+            const float g_exp = exp(g_ptr[0]);
+
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= g_exp;
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        } else {
+            // KDA
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= exp(g_ptr[is]);
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        }
+
+        s_k = simd_sum(s_k);
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        float y = 0.0f;
+
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            ls[j] += k_ptr[is]*d;
+
+            y += ls[j]*q_ptr[is];
+        }
+
+        y = simd_sum(y);
+
+        if (tx == 0) {
+            dst_attn[t*args.ne21*S_v] = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        if (K > 1) {
+            const int target_slot = (int)args.ne22 - 1 - (int)t;
+            if (target_slot >= 0 && target_slot < (int)K) {
+                device float * dst_state = (device float *) (dst) + attn_size + (uint)target_slot * state_size_per_snap + state_out_base;
+                FOR_UNROLL (short j = 0; j < NSG; j++) {
+                    const short is = tx*NSG + j;
+                    dst_state[is] = ls[j];
+                }
+            }
+        }
+    }
+
+    if (K == 1) {
+        device float * dst_state = (device float *) (dst) + attn_size + state_out_base;
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            dst_state[is] = ls[j];
+        }
+    }
+
+#undef S_v
+#undef G
+#undef K
+}
+
+typedef decltype(kernel_gated_delta_net_impl<4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<4>;
+
+#else
+// a simplified version of the above
+// no performance improvement, so keep the above version for now
+
+template<typename T, short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float lsf[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        lsf[j] = s_ptr[is*S_v];
+    }
+
+    thread T * ls = (thread T *) (lsf);
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr  = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr  = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        device const T * qt_ptr = (device const T *) (q_ptr);
+        device const T * kt_ptr = (device const T *) (k_ptr);
+        device const T * gt_ptr = (device const T *) (g_ptr);
+
+        if (G == 1) {
+            *ls *= exp(g_ptr[0]);
+        } else {
+            // KDA
+            *ls *= exp(gt_ptr[tx]);
+        }
+
+        const float s_k = simd_sum(dot(*ls, kt_ptr[tx]));
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        *ls += kt_ptr[tx]*d;
+
+        const float y = simd_sum(dot(*ls, qt_ptr[tx]));
+
+        if (tx == 0) {
+            *dst_attn = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        dst_attn += args.ne21*S_v;
+    }
+
+    device float * dst_state  = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+    device T     * dstt_state = (device T     *) (dst_state);
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = lsf[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<float4, 4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float,  1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float2, 2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float4, 4>;
+#endif
@@ -0,0 +1,595 @@
+#include "common.h"
+
+kernel void kernel_argmax_f32(
+        constant ggml_metal_kargs_argmax & args,
+        device   const char * src0,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    device const float * x_row = (device const float *) ((device const char *) src0 + tgpig * args.nb01);
+
+    float   lmax = -INFINITY;
+    int32_t larg = -1;
+
+    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+        if (x_row[i00] > lmax) {
+            lmax = x_row[i00];
+            larg = i00;
+        }
+    }
+
+    // find the argmax value in the block
+    float max_val = simd_max(lmax);
+    int32_t arg_val = simd_max(select(-1, larg, lmax == max_val));
+
+    device int32_t * dst_i32 = (device int32_t *) dst;
+
+    threadgroup   float * shared_maxval = (threadgroup   float *) shmem;
+    threadgroup int32_t * shared_argmax = (threadgroup int32_t *) shmem + N_SIMDWIDTH;
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            shared_maxval[tiisg] = -INFINITY;
+            shared_argmax[tiisg] = -1;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            shared_maxval[sgitg] = max_val;
+            shared_argmax[sgitg] = arg_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = shared_maxval[tiisg];
+        arg_val = shared_argmax[tiisg];
+
+        float max_val_reduced   = simd_max(max_val);
+        int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced));
+
+        dst_i32[tgpig] = arg_val_reduced;
+
+        return;
+    }
+
+    dst_i32[tgpig] = arg_val;
+}
+
+kernel void kernel_diag_f32(
+        constant ggml_metal_kargs_diag & args,
+        device   const char * src0,
+        device         char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *)(src0 +                i2*args.nb02 + i3*args.nb03);
+    device       float * dst_ptr  = (device       float *)(dst  + i1*args.nb01 + i2*args.nb2  + i3*args.nb3);
+
+    for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}
+
+kernel void kernel_roll_f32(
+    constant ggml_metal_kargs_roll & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *) src0;
+    device       float * dst_ptr  = (device       float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        // apply shifts and wrap around
+        int64_t i00 = i0 - args.s0;
+        int64_t i01 = i1 - args.s1;
+        int64_t i02 = i2 - args.s2;
+        int64_t i03 = i3 - args.s3;
+
+        if (i00 < 0) { i00 += args.ne00; } else if (i00 >= args.ne00) { i00 -= args.ne00; }
+        if (i01 < 0) { i01 += args.ne01; } else if (i01 >= args.ne01) { i01 -= args.ne01; }
+        if (i02 < 0) { i02 += args.ne02; } else if (i02 >= args.ne02) { i02 -= args.ne02; }
+        if (i03 < 0) { i03 += args.ne03; } else if (i03 >= args.ne03) { i03 -= args.ne03; }
+
+        int64_t src_idx = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00 + i00;
+        int64_t dst_idx = i3 *args.ne2 *args.ne1 *args.ne0  + i2 *args.ne1 *args.ne0  + i1 *args.ne0  + i0;
+
+        dst_ptr[dst_idx] = src0_ptr[src_idx];
+    }
+}
+
+template <typename T>
+kernel void kernel_pad_impl(
+    constant ggml_metal_kargs_pad & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t k0 = tgpig.x/args.ne1;
+    const int32_t i1 = tgpig.x - k0*args.ne1;
+
+    const int32_t i03 = i3;
+    const int32_t i02 = i2;
+    const int32_t i01 = i1;
+
+    device const T * src0_ptr = (device const T *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T * dst_ptr  = (device       T *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    for (int32_t l0 = 0; l0 < 1024; l0 += ntg.x) {
+        const int32_t i0 = k0*1024 + tpitg.x + l0;
+        if (i0 >= args.ne0) {
+            break;
+        }
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            dst_ptr[i0] = src0_ptr[i0];
+        } else {
+            dst_ptr[i0] = 0.0f;
+        }
+    }
+}
+
+typedef decltype(kernel_pad_impl<float>) kernel_pad_t;
+
+template [[host_name("kernel_pad_f32")]]   kernel kernel_pad_t kernel_pad_impl<float>;
+template [[host_name("kernel_pad_f32_4")]] kernel kernel_pad_t kernel_pad_impl<float4>;
+
+// TODO: this is slow - optimize
+kernel void kernel_pad_reflect_1d_f32(
+    constant   ggml_metal_kargs_pad_reflect_1d & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3  tgpg[[threadgroups_per_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            if (i0 < args.p0) {
+                dst_ptr[i0] = src0_ptr[args.p0 - i0];
+            } else if (i0 < args.ne0 - args.p1) {
+                dst_ptr[i0] = src0_ptr[i0 - args.p0];
+            } else {
+                dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1];
+            }
+        }
+    }
+}
+
+kernel void kernel_arange_f32(
+    constant   ggml_metal_kargs_arange & args,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    device float * dst_ptr = (device float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_ptr[i0] = args.start + args.step * i0;
+    }
+}
+
+kernel void kernel_timestep_embedding_f32(
+    constant  ggml_metal_kargs_timestep_embedding & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    int i = tgpig.x;
+    device float * embed_data = (device float *)(dst + i*args.nb1);
+
+    int half_ = args.dim / 2;
+    for (int j = tpitg.x; j < half_; j += ntg.x) {
+        float timestep = ((device float *)src0)[i];
+        float freq = (float)exp(-log((float)args.max_period) * j / half_);
+        float arg = timestep * freq;
+        embed_data[j        ] = cos(arg);
+        embed_data[j + half_] = sin(arg);
+    }
+
+    if (args.dim % 2 != 0 && tpitg.x == 0) {
+        embed_data[2 * half_] = 0.f;
+    }
+}
+
+kernel void kernel_opt_step_adamw_f32(
+        constant    ggml_metal_kargs_opt_step_adamw & args,
+        device       float * x,
+        device const float * g,
+        device       float * g_m,
+        device       float * g_v,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
+    const float gi = g[gid];
+    const float gmi = g_m[gid] * beta1 +      gi * (1.0f - beta1);
+    const float gvi = g_v[gid] * beta2 + gi * gi * (1.0f - beta2);
+
+    g_m[gid] = gmi;
+    g_v[gid] = gvi;
+
+    const float mh =      gmi * beta1h;
+    const float vh = sqrt(gvi * beta2h) + eps;
+
+    x[gid] = x[gid] * (1.0f - alpha * wd) - alpha * mh / vh;
+}
+
+kernel void kernel_opt_step_sgd_f32(
+        constant    ggml_metal_kargs_opt_step_sgd & args,
+        device       float * x,
+        device const float * g,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    x[gid] = x[gid] * (1.0f - pars[0] * pars[1]) - pars[0] * g[gid];
+}
+
+template<typename T>
+kernel void kernel_memset(
+        constant ggml_metal_kargs_memset & args,
+        device T * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
+typedef decltype(kernel_memset<int64_t>) kernel_memset_t;
+
+template [[host_name("kernel_memset_i64")]] kernel kernel_memset_t kernel_memset<int64_t>;
+
+constant short FC_count_equal_nsg [[function_constant(FC_COUNT_EQUAL + 0)]];
+
+template<typename T>
+kernel void kernel_count_equal(
+        constant ggml_metal_kargs_count_equal & args,
+        device   const char * src0,
+        device   const char * src1,
+        device   atomic_int * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const short NSG = FC_count_equal_nsg;
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    int sum = 0;
+
+    device const char * base0 = src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03;
+    device const char * base1 = src1 + i1*args.nb11 + i2*args.nb12 + i3*args.nb13;
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        const T v0 = *(device const T *)(base0 + i0*args.nb00);
+        const T v1 = *(device const T *)(base1 + i0*args.nb10);
+        sum += (v0 == v1);
+    }
+
+    sum = simd_sum(sum);
+
+    if (tiisg == 0) {
+        shmem_i32[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        float v = 0.0f;
+        if (tpitg.x < NSG) {
+            v = shmem_i32[tpitg.x];
+        }
+
+        float total = simd_sum(v);
+        if (tpitg.x == 0) {
+            atomic_fetch_add_explicit(dst, (int32_t) total, memory_order_relaxed);
+        }
+    }
+}
+
+typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;
+
+template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;
+
+template <typename T>
+kernel void kernel_snake(
+        constant ggml_metal_kargs_snake & args,
+        device const T     * x,
+        device const float * a,
+        device const float * inv_b,
+        device       T     * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T * args.C) {
+        return;
+    }
+
+    const int   c  = idx / args.T;  // x is [T, C], a / inv_b collapse to [1, C]
+    const float xi = float(x[idx]);
+    const float si = sin(a[c] * xi);
+    dst[idx] = T(xi + si * si * inv_b[c]);
+}
+
+template [[host_name("kernel_snake_f32")]]  kernel void kernel_snake<float>(constant ggml_metal_kargs_snake &, device const float *, device const float *, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_snake_f16")]]  kernel void kernel_snake<half>(constant ggml_metal_kargs_snake &, device const half *, device const float *, device const float *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_snake_bf16")]] kernel void kernel_snake<bfloat>(constant ggml_metal_kargs_snake &, device const bfloat *, device const float *, device const float *, device bfloat *, uint, uint, uint);
+#endif
+
+template<int N>
+kernel void kernel_fwht_f32(
+        constant ggml_metal_kargs_fwht & args,
+        device const float * src,
+        device float * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort3  ntg[[threads_per_threadgroup]]) {
+
+    constexpr int NW = N_SIMDWIDTH;
+    constexpr int NE = N / NW;
+
+    const float scale = 1.0f / sqrt((float) N);
+
+    const int sg_per_tg = ntg.x / NW;
+    const int64_t r = tgpig.x * sg_per_tg + sgitg;
+    if (r >= args.nrows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    const int lane = tiisg;
+
+    float reg[NE];
+    for (int i = 0; i < NE; i++) {
+        reg[i] = src[i*NW + lane]*scale;
+    }
+    for (int i = 1; i < NW; i *= 2) {
+        for (int j = 0; j < NE; j++) {
+            const float val = reg[j];
+            const float val2 = simd_shuffle_xor(val, i);
+            reg[j] = (lane & i) == 0 ? val2 + val : val2 - val;
+        }
+    }
+
+    for (int i = NW; i < N; i *= 2) {
+        const int step = i / NW;
+        for (int j = 0; j < NE; j += (2 * step)) {
+            for (int k = 0; k < step; k++) {
+                const float x = reg[j + k ];
+                const float y = reg[j + k + step];
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+    for (int i = 0; i < NE; i++) {
+        dst[i*NW + lane] = reg[i];
+    }
+}
+
+typedef decltype(kernel_fwht_f32<64>) kernel_fwht_t;
+
+template [[host_name("kernel_fwht_f32_64")]]  kernel kernel_fwht_t kernel_fwht_f32<64>;
+template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_t kernel_fwht_f32<128>;
+template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_t kernel_fwht_f32<256>;
+template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_t kernel_fwht_f32<512>;
+
+kernel void kernel_dsv4_hc_comb_f32(
+        constant ggml_metal_kargs_dsv4_hc_comb & args,
+        device const char * mixes,
+        device const char * scale,
+        device const char * base,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+    constexpr ushort comb_offset = 2*hc;
+
+    const int it = tgpig.x*ntg.y + sgitg;
+    if (it >= args.n_tokens) {
+        return;
+    }
+
+    float scale_lane = 0.0f;
+    if (tiisg == 0) {
+        scale_lane = *(device const float *) (scale + 2*args.nb_s0);
+    }
+    const float scale_comb = simd_shuffle(scale_lane, 0);
+
+    float v = 0.0f;
+    if (tiisg < hc*hc) {
+        v = *(device const float *) (mixes + (comb_offset + tiisg)*args.nb_m0 + it*args.nb_m1)*scale_comb
+          + *(device const float *) (base   + (comb_offset + tiisg)*args.nb_b0);
+    }
+
+    // Softmax across destinations (the four contiguous lanes for each source).
+    float vmax = max(v, simd_shuffle_xor(v, 1));
+    vmax = max(vmax, simd_shuffle_xor(vmax, 2));
+    v = exp(v - vmax);
+
+    float sum = v + simd_shuffle_xor(v, 1);
+    sum += simd_shuffle_xor(sum, 2);
+    v = v/sum + args.eps;
+
+    // Normalize columns: equal destination indices are four lanes apart.
+    sum = v + simd_shuffle_xor(v, 4);
+    sum += simd_shuffle_xor(sum, 8);
+    v /= sum + args.eps;
+
+    for (int i = 1; i < args.n_iter; ++i) {
+        sum = v + simd_shuffle_xor(v, 1);
+        sum += simd_shuffle_xor(sum, 2);
+        v /= sum + args.eps;
+
+        sum = v + simd_shuffle_xor(v, 4);
+        sum += simd_shuffle_xor(sum, 8);
+        v /= sum + args.eps;
+    }
+
+    if (tiisg < hc*hc) {
+        const ushort idst = tiisg & 3;
+        const ushort isrc = tiisg >> 2;
+        *(device float *) (dst + idst*args.nb_d0 + isrc*args.nb_d1 + it*args.nb_d2) = v;
+    }
+}
+
+kernel void kernel_dsv4_hc_pre_f32(
+        constant ggml_metal_kargs_dsv4_hc_pre & args,
+        device const char * x,
+        device const char * weights,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float weight_lane = 0.0f;
+    if (tiisg < hc) {
+        weight_lane = *(device const float *) (weights + tiisg*args.nb_w0 + it*args.nb_w1);
+    }
+
+    float w[hc];
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        w[ih] = simd_shuffle(weight_lane, ih);
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    device const char * xb = x + i0*args.nb_x0 + it*args.nb_x2;
+    float result = 0.0f;
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        result = fma(*(device const float *) (xb + ih*args.nb_x1), w[ih], result);
+    }
+
+    *(device float *) (dst + i0*args.nb_d0 + it*args.nb_d1) = result;
+}
+
+kernel void kernel_dsv4_hc_post_f32(
+        constant ggml_metal_kargs_dsv4_hc_post & args,
+        device const char * x,
+        device const char * residual,
+        device const char * post,
+        device const char * comb,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float coeff_lane = 0.0f;
+    if (tiisg < hc) {
+        coeff_lane = *(device const float *) (post + tiisg*args.nb_p0 + it*args.nb_p1);
+    } else if (tiisg < hc + hc*hc) {
+        const ushort idx  = tiisg - hc;
+        const ushort idst = idx & 3;
+        const ushort isrc = idx >> 2;
+        coeff_lane = *(device const float *) (comb + idst*args.nb_c0 + isrc*args.nb_c1 + it*args.nb_c2);
+    }
+
+    float post_reg[hc];
+    float comb_reg[hc][hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        post_reg[idst] = simd_shuffle(coeff_lane, idst);
+    }
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            comb_reg[isrc][idst] = simd_shuffle(coeff_lane, hc + idst + hc*isrc);
+        }
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    const float xv = *(device const float *) (x + i0*args.nb_x0 + it*args.nb_x1);
+    float result[hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        result[idst] = xv*post_reg[idst];
+    }
+
+    device const char * rb = residual + i0*args.nb_r0 + it*args.nb_r2;
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        const float rv = *(device const float *) (rb + isrc*args.nb_r1);
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            result[idst] = fma(rv, comb_reg[isrc][idst], result[idst]);
+        }
+    }
+
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        *(device float *) (dst + i0*args.nb_d0 + idst*args.nb_d1 + it*args.nb_d2) = result[idst];
+    }
+}
@@ -0,0 +1,853 @@
+#include "common.h"
+#include "dequantize.h"
+
+constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
+constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
+constant short FC_mul_mm_ne12  [[function_constant(FC_MUL_MM + 2)]];
+constant short FC_mul_mm_ne13  [[function_constant(FC_MUL_MM + 3)]];
+constant short FC_mul_mm_r2    [[function_constant(FC_MUL_MM + 4)]];
+constant short FC_mul_mm_r3    [[function_constant(FC_MUL_MM + 5)]];
+
+// each block_q contains 16*nl weights
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    // note: K is dynamic_extent (clamped to the valid range in PHASE 2), since a static
+    //       N_MM_NK_TOTAL K tile would read src1 out of bounds when K % N_MM_NK_TOTAL != 0
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27064
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, static_cast<int>(dynamic_extent), false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        // Clamp the K extent of both operand tensors to the remaining valid K range so
+        // the dynamic-K op never reads past the K extent of src1 (or the staged A tile).
+        const int kExt = min(N_MM_NK_TOTAL, K - loop_k);
+
+        auto tAv = tensor(sa, dextents<int32_t, 2>(kExt, NRA), array<int, 2>({1, N_MM_NK_TOTAL}));
+        auto tBv = tensor(ptrB + loop_k + rb * strideB, dextents<int32_t, 2>(kExt, N - rb), array<int, 2>({1, strideB}));
+
+        mm.run(tBv, tAv, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+    }
+
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
+        // if no bounds checks on the output are needed, we can directly write to device memory
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (sgitg == 0) {
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*NR0);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < nr0/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < nr0; i++) {
+                    *(D + i) = *(C + i);
+                }
+            }
+        }
+    }
+}
+
+#endif // GGML_METAL_HAS_TENSOR
+
+template<short ne20> // n_expert_used
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src2,
+        device        char * htpe,
+        device        char * hids,
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
+
+    uint32_t n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
+
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
+
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
+}
+
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
+
+template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * htpe,
+        device const char * hids,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
+
+    const int32_t neh1 = tpe_u32[im];
+
+    if (r1 >= neh1) {
+        return;
+    }
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifndef GGML_METAL_HAS_TENSOR
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
+    }
+
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+    }
+#endif
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < nr0/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
+        }
+
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
+            *(D + i) = *(C + i);
+        }
+    }
+}
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_q1_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_tq2_0_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_q1_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_tq2_0_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  half, half2x4>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_tq2_0_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_tq2_0_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  half, half2x4>;
@@ -0,0 +1,308 @@
+#include "common.h"
+
+// F == 1 : norm (no fuse)
+// F == 2 : norm + mul
+// F == 3 : norm + mul + add
+template <typename T, short F>
+kernel void kernel_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    T sumft(0.0f);
+
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumft += x[i00];
+    }
+    sumf = dot(sumft, T(1.0f));
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+
+    sumf = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        y[i00] = x[i00] - mean;
+        sumf += dot(y[i00], y[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (y[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (y[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (y[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_norm_fuse_impl<float4, 1>) kernel_norm_fuse_t;
+
+template [[host_name("kernel_norm_f32")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_norm_mul_f32")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_norm_mul_add_f32")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_norm_f32_4")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_norm_mul_f32_4")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_norm_mul_add_f32_4")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 3>;
+
+// F == 1 : rms_norm (no fuse)
+// F == 2 : rms_norm + mul
+// F == 3 : rms_norm + mul + add
+template <typename T, short F>
+kernel void kernel_rms_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (x[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (x[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (x[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_rms_norm_fuse_impl<float4, 1>) kernel_rms_norm_fuse_t;
+
+template [[host_name("kernel_rms_norm_f32")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_rms_norm_mul_f32")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_rms_norm_f32_4")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_rms_norm_mul_f32_4")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32_4")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 3>;
+
+template <typename T0, typename T>
+kernel void kernel_l2_norm_impl(
+        constant ggml_metal_kargs_l2_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const T0 * x = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T  * y = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float scale = 1.0f/max(sqrt(sumf), args.eps);
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+typedef decltype(kernel_l2_norm_impl<float, float>) kernel_l2_norm_t;
+
+template [[host_name("kernel_l2_norm_f32_f32")]]   kernel kernel_l2_norm_t kernel_l2_norm_impl<float,  float>;
+template [[host_name("kernel_l2_norm_f32_f32_4")]] kernel kernel_l2_norm_t kernel_l2_norm_impl<float4, float4>;
+
+kernel void kernel_group_norm_f32(
+        constant ggml_metal_kargs_group_norm & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    const int64_t ne = args.ne00*args.ne01*args.ne02;
+    const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.ngrp - 1) / args.ngrp);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}
@@ -0,0 +1,148 @@
+#include "common.h"
+
+kernel void kernel_pool_2d_max_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+
+    float res = -INFINITY;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            res = MAX(res, i_ptr[i * args.IW + j]);
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+kernel void kernel_pool_2d_avg_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+    // const float scale = 1. / ((eh - bh) * (ew - bw));
+    const float scale = 1. / (args.k0 * args.k1);
+
+    float res = 0;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            float cur = i_ptr[i * args.IW + j];
+            res += cur * scale;
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+
+kernel void kernel_pool_1d_max_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = -INFINITY;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        int j = base + ki;
+        if (j < 0 || j >= args.IW){
+            continue;
+        }
+        float v = src[src_off + j];
+        acc = max(acc, v);
+    }
+
+    dst[dst_off + ow] = acc;
+}
+
+kernel void kernel_pool_1d_avg_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = 0.0f;
+    int   cnt = 0;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        const int j = base + ki;
+        if (j < 0 || j >= args.IW) {
+            continue;
+        }
+        acc += src[src_off + j];
+        cnt += 1;
+    }
+
+    dst[dst_off + ow] = (cnt > 0) ? (acc / (float)cnt) : 0.0f;
+}
@@ -0,0 +1,262 @@
+#pragma once
+
+#include "common.h"
+
+void quantize_q1_0(device const float * src, device block_q1_0 & dst) {
+    float sum_abs = 0.0f;
+    for (int j = 0; j < QK1_0; j++) {
+        sum_abs += fabs(src[j]);
+    }
+    dst.d = sum_abs / QK1_0;
+
+    for (int j = 0; j < QK1_0 / 8; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK1_0; j++) {
+        if (src[j] >= 0.0f) {
+            dst.qs[j / 8] |= (1 << (j % 8));
+        }
+    }
+}
+
+void quantize_q2_0(device const float * src, device block_q2_0 & dst) {
+    float amax = 0.0f;
+    for (int j = 0; j < QK2_0; j++) {
+        float a = fabs(src[j]);
+        if (a > amax) amax = a;
+    }
+    const float d = amax;
+    dst.d = d;
+
+    const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+    for (int j = 0; j < QK2_0 / 4; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK2_0; j++) {
+        int q = (int)round(src[j] * id) + 1;
+        q = max(0, min(3, q));
+        dst.qs[j / 4] |= (q << (2 * (j % 4)));
+    }
+}
+
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[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));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[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));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[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;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[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 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
+void quantize_tq2_0(device const float * src, device block_tq2_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK_K; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = (half) d;
+
+    for (int j = 0; j < QK_K/4; j += 32) {
+        for (int m = 0; m < 32; ++m) {
+            uint8_t q = 0;
+            for (int n = 0; n < 4; ++n) {
+                // -1, 0, 1 -> 0, 1, 2
+                int xi = (int)round(src[m + n*32] * id) + 1;
+                q += (uint8_t)((xi & 3) << (2*n));
+            }
+            dst.qs[j + m] = q;
+        }
+        src += 4*32;
+    }
+}
@@ -0,0 +1,435 @@
+#include "common.h"
+#include "dequantize.h"
+#include "quantize.h"
+
+template<typename T0, typename T1>
+kernel void kernel_cpy_t_t(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.ne00;) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        dst_data[i00] = (T1) src[0];
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_t_t<float, float>) kernel_cpy_t;
+
+template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   float>;
+template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
+template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
+template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
+#endif
+template [[host_name("kernel_cpy_f16_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    float>;
+template [[host_name("kernel_cpy_f16_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  float>;
+template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  bfloat>;
+#endif
+
+template<short QK,
+         typename block_q,
+         void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_cpy_f32_q(
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK;
+
+    device block_q * dst_data = (device block_q *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        device const float * src = (device const float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + (i00*QK)*args.nb00);
+
+        quantize_func(src, dst_data[i00]);
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_f32_q<QK8_0,  block_q8_0,  quantize_q8_0>)  cpy_f_q_t;
+
+template [[host_name("kernel_cpy_f32_q8_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK8_0,  block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_cpy_f32_q1_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK1_0,  block_q1_0,   quantize_q1_0>;
+template [[host_name("kernel_cpy_f32_q2_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK2_0,  block_q2_0,   quantize_q2_0>;
+template [[host_name("kernel_cpy_f32_q4_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_0,  block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_cpy_f32_q4_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_1,  block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_cpy_f32_q5_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_0,  block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_cpy_f32_q5_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_1,  block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_cpy_f32_iq4_nl")]] kernel cpy_f_q_t kernel_cpy_f32_q<QK4_NL, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_cpy_f32_tq2_0")]]  kernel cpy_f_q_t kernel_cpy_f32_q<QK_K,   block_tq2_0,  quantize_tq2_0>;
+
+template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_cpy_q_f32(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T4x4    * dst_data = (device       T4x4    *)(dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        T4x4 temp;
+        dequantize_func(src_data + i00/nl, i00%nl, temp);
+        dst_data[i00] = temp;
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>) cpy_q_f_t;
+
+template [[host_name("kernel_cpy_q1_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_tq2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_tq2_0, QK_NL, dequantize_tq2_0>;
+
+template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_tq2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_tq2_0, QK_NL, dequantize_tq2_0>;
+
+template<typename T>
+kernel void kernel_concat(
+        constant ggml_metal_kargs_concat & args,
+        device  const char * src0,
+        device  const char * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = ntg.y == 1 ? tgpig.x : tgpig.x*ntg.y + tpitg.y;
+
+    if (i1 >= args.ne1) {
+        return;
+    }
+
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        device const T * x;
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const T *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
+        } else {
+            x = (device const T *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
+        }
+
+        device T * y = (device T *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+        *y = *x;
+    }
+}
+
+typedef decltype(kernel_concat<float>) kernel_concat_t;
+
+template [[host_name("kernel_concat_f32")]]  kernel kernel_concat_t kernel_concat<float>;
+template [[host_name("kernel_concat_f16")]]  kernel kernel_concat_t kernel_concat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_concat_bf16")]] kernel kernel_concat_t kernel_concat<bfloat>;
+#endif
+template [[host_name("kernel_concat_i8")]]   kernel kernel_concat_t kernel_concat<char>;
+template [[host_name("kernel_concat_i16")]]  kernel kernel_concat_t kernel_concat<short>;
+template [[host_name("kernel_concat_i32")]]  kernel kernel_concat_t kernel_concat<int>;
+template [[host_name("kernel_concat_i64")]]  kernel kernel_concat_t kernel_concat<long>;
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg  [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (device const block_q *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (device      float4x4 *) ((      device char *) dst  + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        float4x4 temp;
+        dequantize_func(psrc + ind/nl, ind%nl, temp);
+        pdst[ind] = temp;
+
+        break;
+    }
+}
+
+template<typename T0, typename T>
+kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (const device T0 *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (      device T  *) ((      device char *)  dst + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        pdst[ind] = psrc[ind];
+
+        break;
+    }
+}
+
+typedef decltype(kernel_get_rows_f<float, float>) get_rows_f_t;
+
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float, float>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half,  float>;
+template [[host_name("kernel_get_rows_i32")]]  kernel get_rows_f_t kernel_get_rows_f<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat, float>;
+#endif
+
+typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
+
+template [[host_name("kernel_get_rows_q1_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q1_0,    8, dequantize_q1_0>;
+template [[host_name("kernel_get_rows_q2_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_0,    4, dequantize_q2_0>;
+template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
+template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+template [[host_name("kernel_get_rows_tq2_0")]]   kernel get_rows_q_t kernel_get_rows_q<block_tq2_0,   QK_NL, dequantize_tq2_0>;
+
+template<typename TS, typename TI, short QK, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + QK*ind, dst_row[ind]);
+    }
+}
+
+template<typename TS, typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename TS, typename TI, typename TD>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device TD * dst_row = (      device TD *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS * src_row = (const device TS *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (TD) src_row[ind];
+    }
+}
+
+typedef decltype(kernel_set_rows_f<float, int64_t, float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32_i64_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, float>;
+template [[host_name("kernel_set_rows_f32_i32_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, float>;
+template [[host_name("kernel_set_rows_f32_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, half>;
+template [[host_name("kernel_set_rows_f32_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_f32_i64_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_f32_i32_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int32_t, bfloat>;
+#endif
+
+template [[host_name("kernel_set_rows_f16_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int64_t, half>;
+template [[host_name("kernel_set_rows_f16_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_bf16_i64_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_bf16_i32_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int32_t, bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<float, int64_t, block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_f32_i64_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i32_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i32_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i32_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i64_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i32_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i64_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i32_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i64_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_set_rows_f32_i32_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_iq4_nl, quantize_iq4_nl>;
+
+typedef decltype(kernel_set_rows_q<float, int64_t, QK_K, block_tq2_0, quantize_tq2_0>) set_rows_qK_t;
+
+template [[host_name("kernel_set_rows_f32_i64_tq2_0")]]  kernel set_rows_qK_t kernel_set_rows_q<float, int64_t, QK_K, block_tq2_0, quantize_tq2_0>;
+template [[host_name("kernel_set_rows_f32_i32_tq2_0")]]  kernel set_rows_qK_t kernel_set_rows_q<float, int32_t, QK_K, block_tq2_0, quantize_tq2_0>;
+
@@ -0,0 +1,228 @@
+#include "common.h"
+
+kernel void kernel_op_sum_f32(
+        constant ggml_metal_kargs_sum & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    if (args.np == 0) {
+        return;
+    }
+
+    // TODO: become function constant
+    const uint nsg = (ntg.x + 31) / 32;
+
+    float sumf = 0;
+
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+        sumf += src0[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float total = 0;
+
+    if (sgitg == 0) {
+        float v = 0;
+
+        if (tpitg.x < nsg) {
+            v = shmem_f32[tpitg.x];
+        }
+
+        total = simd_sum(v);
+
+        if (tpitg.x == 0) {
+            dst[0] = total;
+        }
+    }
+}
+
+constant short FC_sum_rows_op [[function_constant(FC_SUM_ROWS + 0)]];
+
+template <typename T0, typename T>
+kernel void kernel_sum_rows_impl(
+        constant ggml_metal_kargs_sum_rows & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_sum_rows_op
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    threadgroup T0 * shmem_t = (threadgroup T0 *) shmem;
+
+    if (sgitg == 0) {
+        shmem_t[tiisg] = 0.0f;
+    }
+
+    device const T0 * src_row = (device const T0 *) (src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T  * dst_row = (device       T  *) (dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    T0 sumf = T0(0.0f);
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_t[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_t[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        if (FC_OP == OP_SUM_ROWS_NUM_MEAN) {
+            if (is_same<float4, T0>::value) {
+                dst_row[0] = sum(sumf) / (4*args.ne00);
+            } else {
+                dst_row[0] = sum(sumf) / args.ne00;
+            }
+        } else {
+            dst_row[0] = sum(sumf);
+        }
+    }
+
+#undef FC_OP
+}
+
+typedef decltype(kernel_sum_rows_impl<float, float>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows_f32_f32")]]   kernel kernel_sum_rows_t kernel_sum_rows_impl<float,  float>;
+template [[host_name("kernel_sum_rows_f32_f32_4")]] kernel kernel_sum_rows_t kernel_sum_rows_impl<float4, float>;
+
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;
@@ -0,0 +1,333 @@
+#include "common.h"
+
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+constant bool FC_rope_is_back   [[function_constant(FC_ROPE + 1)]];
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+    if (FC_rope_is_back) {
+        *sin_theta *= -1.0f;
+    }
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
+
+template<typename T>
+kernel void kernel_rope_norm(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            if (args.inplace) {
+                continue;
+            }
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_neox(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            if (args.inplace) {
+                continue;
+            }
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            if (args.inplace) {
+                continue;
+            }
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
+typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
+
+template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
+template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
+
+template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
+template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
+
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;
@@ -0,0 +1,223 @@
+#include "common.h"
+
+template<typename T>
+kernel void kernel_soft_max(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float * psrc2 = src2 != src0 ? (device const float *) (src2)                                                 : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+
+    // find the max value in the block
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
+        lsum += exp_psrc0;
+        pdst[i00] = exp_psrc0;
+    }
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        pdst[i00] *= inv_sum;
+    }
+}
+
+template<typename T>
+kernel void kernel_soft_max_4(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float *  psrc2 = src2 != src0 ? (device const float * ) (src2)                                                 : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
+    }
+
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        pdst4[i00] *= inv_sum;
+    }
+}
+
+typedef decltype(kernel_soft_max<float>)    kernel_soft_max_t;
+typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
+
+template [[host_name("kernel_soft_max_f16")]]   kernel kernel_soft_max_t   kernel_soft_max<half>;
+template [[host_name("kernel_soft_max_f32")]]   kernel kernel_soft_max_t   kernel_soft_max<float>;
+template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
+template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;
@@ -0,0 +1,75 @@
+#include "common.h"
+
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}
@@ -0,0 +1,287 @@
+#include "common.h"
+
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+kernel void kernel_ssm_conv_f32_f32(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+    device       float  * x = (device       float  *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
+kernel void kernel_ssm_scan_f32(
+        constant ggml_metal_kargs_ssm_scan & args,
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgptg[[simdgroups_per_threadgroup]],
+        uint3    tgpg[[threadgroups_per_grid]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
+
+    const int32_t i0 = tpitg.x;
+    const int32_t i1 = tgpig.x;
+    const int32_t ir = tgpig.y; // current head
+    const int32_t i3 = tgpig.z; // current seq
+
+    const int32_t nc  = args.d_state;
+    const int32_t nr  = args.d_inner;
+    const int32_t nh  = args.n_head;
+    const int32_t ng  = args.n_group;
+    const int32_t n_t = args.n_seq_tokens;
+    const int32_t n_s = args.n_seqs;
+    const int32_t K   = args.K;
+
+    const int32_t s_off = args.s_off;
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    const int32_t i = i0 + i1*nc;
+    const int32_t g = ir / (nh / ng); // repeat_interleave
+
+    float s0 = s0_buff[i];
+    float s  = 0.0f;
+
+    device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); // {ne30, nh}
+
+    const float A0 = A[i0%args.ne30];
+
+    device const float * x  = (device const float *)((device const char *) src1 + i1*args.nb10  + ir*args.nb11 + i3*args.nb13); // {dim, nh, nt, ns}
+    device const float * dt = (device const float *)((device const char *) src2 + ir*args.nb20  + i3*args.nb22);                // {nh, nt, ns}
+    device const float * B  = (device const float *)((device const char *) src4 +  g*args.nb41  + i3*args.nb43);                // {d_state, ng, nt, ns}
+    device const float * C  = (device const float *)((device const char *) src5 +  g*args.nb51  + i3*args.nb53);                // {d_state, ng, nt, ns}
+
+    device float * y = dst + (i1 + ir*(nr) + i3*(n_t*nh*nr)); // {dim, nh, nt, ns}
+
+    for (int i2 = 0; i2 < n_t; i2 += sgptg) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
+            const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
+
+            s = (s0 * dA) + (B[i0] * x_dt);
+
+            const float sumf = simd_sum(s * C[i0]);
+
+            if (tiisg == 0) {
+                shared_sums[t*NW + sgitg] = sumf;
+            }
+
+            // recurse
+            s0 = s;
+
+            const int32_t slot = n_t - 1 - (i2 + t);
+            if (slot > 0 && slot < K) {
+                device float * s_snapshot = (device float *) ((device char *) s_buff + (int64_t) slot*n_s*args.nb03);
+                s_snapshot[i] = s;
+            }
+
+            B  += args.ns42;
+            C  += args.ns52;
+        }
+
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
+
+        if (tiisg == 0 && i2 + sgitg < n_t) {
+            y[sgitg*nh*nr] = sumf;
+        }
+
+        y += sgptg*nh*nr;
+    }
+
+    s_buff[i] = s;
+}
@@ -0,0 +1,69 @@
+#include "common.h"
+
+template<uint32_t ttype>
+bool _ggml_vec_tri_cmp(const int i, const int r);
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER */ 3>(const int i, const int r) {
+    return i < r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER_DIAG */ 2>(const int i, const int r) {
+    return i <= r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER */ 1>(const int i, const int r) {
+    return i > r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER_DIAG */ 0>(const int i, const int r) {
+    return i >= r;
+}
+
+template<typename T, int ttype>
+kernel void kernel_tri(
+        constant ggml_metal_kargs_tri & args,
+        device const char * src0,
+        device const char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    device const T * src_row = (device const T *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T * dst_row = (device       T *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    // Each thread is a single element of the row if ne00 < max threads per
+    // threadgroup, so this will loop once for each index that this thread is
+    // responsible for
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        // Use the comparison as a mask for branchless
+        dst_row[i0] = static_cast<T>(_ggml_vec_tri_cmp<ttype>(i0, i1)) * src_row[i0];
+    }
+}
+
+typedef decltype(kernel_tri<float, 0>) kernel_tri_t;
+
+template [[host_name("kernel_tri_f32_0")]] kernel kernel_tri_t kernel_tri<float, 0>;
+template [[host_name("kernel_tri_f32_1")]] kernel kernel_tri_t kernel_tri<float, 1>;
+template [[host_name("kernel_tri_f32_2")]] kernel kernel_tri_t kernel_tri<float, 2>;
+template [[host_name("kernel_tri_f32_3")]] kernel kernel_tri_t kernel_tri<float, 3>;
+template [[host_name("kernel_tri_f16_0")]] kernel kernel_tri_t kernel_tri<half, 0>;
+template [[host_name("kernel_tri_f16_1")]] kernel kernel_tri_t kernel_tri<half, 1>;
+template [[host_name("kernel_tri_f16_2")]] kernel kernel_tri_t kernel_tri<half, 2>;
+template [[host_name("kernel_tri_f16_3")]] kernel kernel_tri_t kernel_tri<half, 3>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_tri_bf16_0")]] kernel kernel_tri_t kernel_tri<bfloat, 0>;
+template [[host_name("kernel_tri_bf16_1")]] kernel kernel_tri_t kernel_tri<bfloat, 1>;
+template [[host_name("kernel_tri_bf16_2")]] kernel kernel_tri_t kernel_tri<bfloat, 2>;
+template [[host_name("kernel_tri_bf16_3")]] kernel kernel_tri_t kernel_tri<bfloat, 3>;
+#endif
@@ -0,0 +1,374 @@
+#include "common.h"
+
+constant short FC_unary_op [[function_constant(FC_UNARY + 0)]];
+constant bool  FC_unary_cnt[[function_constant(FC_UNARY + 1)]];
+
+template <typename T0, typename T, typename TC>
+kernel void kernel_unary_impl(
+        constant ggml_metal_kargs_unary & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_unary_op
+#define FC_CNT FC_unary_cnt
+
+    device const T0 * src0_ptr;
+    device       T  * dst_ptr;
+
+    int i0;
+
+    if (FC_CNT) {
+        i0 = tgpig.x;
+
+        src0_ptr = (device const T0 *) (src0);
+        dst_ptr  = (device       T  *) (dst);
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int k0  = tgpig.x/args.ne01;
+        const int i01 = tgpig.x - k0*args.ne01;
+
+        i0 = k0*ntg.x + tpitg.x;
+
+        src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+        dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1 );
+    }
+
+    {
+        //threadgroup_barrier(mem_flags::mem_none);
+
+        if (!FC_CNT) {
+            if (i0 >= args.ne0) {
+                return;
+            }
+        }
+
+        const TC x = (TC) src0_ptr[i0];
+
+        if (FC_OP == OP_UNARY_NUM_SCALE) {
+            dst_ptr[i0] = (T) (args.scale * x + args.bias);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FILL) {
+            dst_ptr[i0] = (T) args.val;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CLAMP) {
+            dst_ptr[i0] = (T) clamp(x, args.min, args.max);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQR) {
+            dst_ptr[i0] = (T) (x * x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQRT) {
+            dst_ptr[i0] = (T) sqrt(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIN) {
+            dst_ptr[i0] = (T) sin(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_COS) {
+            dst_ptr[i0] = (T) cos(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LOG) {
+            dst_ptr[i0] = (T) log(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LEAKY_RELU) {
+            dst_ptr[i0] = (T) (TC(x > 0)*x + TC(x <= 0)*(x * args.slope));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TANH) {
+            dst_ptr[i0] = (T) precise::tanh(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_RELU) {
+            dst_ptr[i0] = (T) fmax(0, x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIGMOID) {
+            dst_ptr[i0] = (T) (1 / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + precise::tanh(SQRT_2_OVER_PI*x*(1 + GELU_COEF_A*x*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_ERF) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + erf_approx(SQRT_2_INV*x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_QUICK) {
+            dst_ptr[i0] = (T) (x * (1/(1 + exp(GELU_QUICK_COEF*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SILU) {
+            dst_ptr[i0] = (T) (x / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ELU) {
+            dst_ptr[i0] = (T) elu_approx(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_NEG) {
+            dst_ptr[i0] = (T) -x;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ABS) {
+            dst_ptr[i0] = (T) fabs(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SGN) {
+            dst_ptr[i0] = T(x > 0) - T(x < 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_STEP) {
+            dst_ptr[i0] = T(x > 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSWISH) {
+            dst_ptr[i0] = (T) (x * fmax(0, fmin(1, x/6 + 0.5)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSIGMOID) {
+            dst_ptr[i0] = (T) fmax(0, fmin(1, x/6 + 0.5));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXP) {
+            dst_ptr[i0] = (T) exp(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SOFTPLUS) {
+            dst_ptr[i0] = (T) select(log(1 + exp(x)), x, x > 20);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXPM1) {
+            // TODO: precise implementation
+            dst_ptr[i0] = (T) (exp(x) - 1);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_XIELU) {
+            const TC xi      = x;
+            const TC gate    = TC(xi > TC(0.0f));
+            const TC clamped = fmin(xi, TC(args.val));
+            const TC y_pos   = TC(args.scale) * xi * xi + TC(args.bias) * xi;
+            const TC y_neg   = (exp(clamped) - TC(1.0f) - xi) * TC(args.slope) + TC(args.bias) * xi;
+            dst_ptr[i0] = (T) (gate * y_pos + (TC(1.0f) - gate) * y_neg);
+        }
+    }
+
+#undef FC_OP
+#undef FC_CNT
+}
+
+typedef decltype(kernel_unary_impl<float, float, float>) kernel_unary_t;
+
+template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float,  float>;
+template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4, float4>;
+template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
+template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
+
+kernel void kernel_silu_back_f32(
+        constant ggml_metal_kargs_silu_back & args,
+        device const float * dy,
+        device const float * x,
+        device       float * dx,
+        uint gid [[thread_position_in_grid]]) {
+    if (gid >= args.ne) {
+        return;
+    }
+
+    const float s = 1.0f / (1.0f + exp(-x[gid]));
+    dx[gid] = dy[gid] * s * (1.0f + x[gid] * (1.0f - s));
+}
+
+template<typename T>
+kernel void kernel_reglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = (T)(x0*x1*(x0 > 0.0f));
+    }
+}
+
+typedef decltype(kernel_reglu<float>) kernel_reglu_t;
+
+template [[host_name("kernel_reglu_f32")]] kernel kernel_reglu_t kernel_reglu<float>;
+template [[host_name("kernel_reglu_f16")]] kernel kernel_reglu_t kernel_reglu<half>;
+
+template<typename T>
+kernel void kernel_geglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = (T)(gelu*x1);
+    }
+}
+
+typedef decltype(kernel_geglu<float>) kernel_geglu_t;
+
+template [[host_name("kernel_geglu_f32")]] kernel kernel_geglu_t kernel_geglu<float>;
+template [[host_name("kernel_geglu_f16")]] kernel kernel_geglu_t kernel_geglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = (T)(silu*x1);
+    }
+}
+
+typedef decltype(kernel_swiglu<float>) kernel_swiglu_t;
+
+template [[host_name("kernel_swiglu_f32")]] kernel kernel_swiglu_t kernel_swiglu<float>;
+template [[host_name("kernel_swiglu_f16")]] kernel kernel_swiglu_t kernel_swiglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu_oai(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, args.limit);
+        x1 = max(min(x1, args.limit), -args.limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * args.alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = (T)out_glu;
+    }
+}
+
+typedef decltype(kernel_swiglu_oai<float>) kernel_swiglu_oai_t;
+
+template [[host_name("kernel_swiglu_oai_f32")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<float>;
+template [[host_name("kernel_swiglu_oai_f16")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<half>;
+
+template<typename T>
+kernel void kernel_geglu_erf(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = (T)(gelu_erf*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_erf<float>) kernel_geglu_erf_t;
+
+template [[host_name("kernel_geglu_erf_f32")]] kernel kernel_geglu_erf_t kernel_geglu_erf<float>;
+template [[host_name("kernel_geglu_erf_f16")]] kernel kernel_geglu_erf_t kernel_geglu_erf<half>;
+
+template<typename T>
+kernel void kernel_geglu_quick(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = (T)(gelu_quick*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_quick<float>) kernel_geglu_quick_t;
+
+template [[host_name("kernel_geglu_quick_f32")]] kernel kernel_geglu_quick_t kernel_geglu_quick<float>;
+template [[host_name("kernel_geglu_quick_f16")]] kernel kernel_geglu_quick_t kernel_geglu_quick<half>;
@@ -0,0 +1,179 @@
+#include "common.h"
+
+constant bool FC_upscale_aa [[function_constant(FC_UPSCALE + 0)]];
+
+kernel void kernel_upscale_nearest_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3/args.sf3;
+    const int64_t i02 = i2/args.sf2;
+    const int64_t i01 = i1/args.sf1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int64_t i00 = i0/args.sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1  +  i0*args.nb0);
+
+        dst_ptr[0] = src0_ptr[0];
+    }
+}
+
+static inline float bilinear_tri(float x) {
+    return MAX(0.0f, 1.0f - fabs(x));
+}
+
+kernel void kernel_upscale_bilinear_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01  = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01  = MAX(0, MIN(args.ne01 - 1, (int64_t)floor(f01)));
+    const int64_t i01p = MAX(0, MIN(args.ne01 - 1, i01 + 1));
+    const float   fd1  = MAX(0.0f, MIN(1.0f, f01 - (float)i01));
+
+    src0 += i03*args.nb03 + i02*args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1);
+
+    if (FC_upscale_aa) {
+        const float support0  = MAX(1.0f, 1.0f / args.sf0);
+        const float invscale0 = 1.0f / support0;
+        const float support1  = MAX(1.0f, 1.0f / args.sf1);
+        const float invscale1 = 1.0f / support1;
+
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+
+            int64_t x_min = MAX((int64_t)0, (int64_t)floor(f00 - support0 + args.poffs));
+            int64_t x_max = MIN(args.ne00,  (int64_t)ceil (f00 + support0 + args.poffs));
+
+            int64_t y_min = MAX((int64_t)0, (int64_t)floor(f01 - support1 + args.poffs));
+            int64_t y_max = MIN(args.ne01,  (int64_t)ceil (f01 + support1 + args.poffs));
+
+            float sum = 0.0f;
+            float wsum = 0.0f;
+
+            for (int64_t sy = y_min; sy < y_max; ++sy) {
+                const float wy = MAX(0.0f, 1.0f - fabs((float)sy - f01) * invscale1);
+                for (int64_t sx = x_min; sx < x_max; ++sx) {
+                    const float wx = MAX(0.0f, 1.0f - fabs((float)sx - f00) * invscale0);
+                    const float w  = wx * wy;
+                    device const float * src_ptr = (device const float *)(src0 + sy*args.nb01 + sx*args.nb00);
+                    sum  += (*src_ptr) * w;
+                    wsum += w;
+                }
+            }
+
+            const float v = (wsum > 0.0f) ? (sum / wsum) : 0.0f;
+            dst_ptr[i0] = v;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float   f00  = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+            const int64_t i00  = MAX(0, MIN(args.ne00 - 1, (int64_t)floor(f00)));
+            const int64_t i00p = MAX(0, MIN(args.ne00 - 1, i00 + 1));
+            const float   fd0  = MAX(0.0f, MIN(1.0f, f00 - (float)i00));
+
+            device const float * src00 = (device const float *)(src0 + i01*args.nb01  + i00*args.nb00);
+            device const float * src10 = (device const float *)(src0 + i01*args.nb01  + i00p*args.nb00);
+            device const float * src01 = (device const float *)(src0 + i01p*args.nb01 + i00*args.nb00);
+            device const float * src11 = (device const float *)(src0 + i01p*args.nb01 + i00p*args.nb00);
+
+            const float v =
+                (*src00) * (1.0f - fd0) * (1.0f - fd1) +
+                (*src10) * fd0          * (1.0f - fd1) +
+                (*src01) * (1.0f - fd0) * fd1 +
+                (*src11) * fd0          * fd1;
+
+            dst_ptr[i0] = v;
+        }
+    }
+}
+
+static inline float bicubic_weight1(float x) {
+    const float a = -0.75f;
+    return ((a + 2) * x - (a + 3)) * x * x + 1;
+}
+
+static inline float bicubic_weight2(float x) {
+    const float a = -0.75f;
+    return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a;
+}
+
+kernel void kernel_upscale_bicubic_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01 = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01 = (int64_t)floor(f01);
+    const float   fd1 = f01 - (float)i01;
+
+    const float w_y0 = bicubic_weight2(fd1 + 1.0f);
+    const float w_y1 = bicubic_weight1(fd1);
+    const float w_y2 = bicubic_weight1(1.0f - fd1);
+    const float w_y3 = bicubic_weight2(2.0f - fd1);
+
+    const device char * src_slice = src0 + i03 * args.nb03 + i02 * args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3 * args.nb3 + i2 * args.nb2 + i1 * args.nb1);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const float   f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+        const int64_t i00 = (int64_t)floor(f00);
+        const float   fd0 = f00 - (float)i00;
+
+        const float w_x0 = bicubic_weight2(fd0 + 1.0f);
+        const float w_x1 = bicubic_weight1(fd0);
+        const float w_x2 = bicubic_weight1(1.0f - fd0);
+        const float w_x3 = bicubic_weight2(2.0f - fd0);
+
+        float sum = 0.0f;
+
+        for (int dy = -1; dy <= 2; ++dy) {
+            const int64_t iy = MAX(0, MIN(args.ne01 - 1, i01 + dy));
+            const float wy = (dy == -1) ? w_y0 : (dy == 0) ? w_y1 : (dy == 1) ? w_y2 : w_y3;
+
+            for (int dx = -1; dx <= 2; ++dx) {
+                const int64_t ix = MAX(0, MIN(args.ne00 - 1, i00 + dx));
+                const float wx = (dx == -1) ? w_x0 : (dx == 0) ? w_x1 : (dx == 1) ? w_x2 : w_x3;
+
+                device const float * src_ptr = (device const float *)(src_slice + iy * args.nb01 + ix * args.nb00);
+                sum += (*src_ptr) * wx * wy;
+            }
+        }
+
+        dst_ptr[i0] = sum;
+    }
+}
@@ -0,0 +1,179 @@
+#include "common.h"
+
+kernel void kernel_rwkv_wkv6_f32(
+    device const float * k,
+    device const float * v,
+    device const float * r,
+    device const float * tf,
+    device const float * td,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _k[head_size];
+    threadgroup float _r[head_size];
+    threadgroup float _tf[head_size];
+    threadgroup float _td[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    _tf[tid] = tf[head_id * head_size + tid];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0;
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 tf_vec = float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            float4 td_vec = float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            float4 temp = tf_vec * kv + s_vec;
+            y += dot(r_vec, temp);
+
+            s_vec = s_vec * td_vec + kv;
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
+
+kernel void kernel_rwkv_wkv7_f32(
+    device const float * r,
+    device const float * w,
+    device const float * k,
+    device const float * v,
+    device const float * a,
+    device const float * b,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _r[head_size];
+    threadgroup float _w[head_size];
+    threadgroup float _k[head_size];
+    threadgroup float _a[head_size];
+    threadgroup float _b[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + tid * head_size + i];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0, sa = 0.0;
+
+        float4 sa_vec(0.0);
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 a_vec = float4(_a[j], _a[j+1], _a[j+2], _a[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+            sa_vec += a_vec * s_vec;
+        }
+        sa = sa_vec[0] + sa_vec[1] + sa_vec[2] + sa_vec[3];
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 w_vec = float4(_w[j], _w[j+1], _w[j+2], _w[j+3]);
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 b_vec = float4(_b[j], _b[j+1], _b[j+2], _b[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            s_vec = s_vec * w_vec + kv + sa * b_vec;
+            y += dot(s_vec, r_vec);
+
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + tid * head_size + i] = state[i];
+    }
+}
@@ -63,6 +63,7 @@ endfunction()
 set(GGML_OPENCL_KERNELS
    add
    add_id
+    moe_add_id_glu
    argsort
    tri
    fill
@@ -202,6 +203,7 @@ set(GGML_OPENCL_KERNELS
    sqr
    sqrt
    ssm_conv
+    ssm_scan
    gated_delta_net
    sub
    sum_rows
@@ -577,11 +577,19 @@ struct ggml_backend_opencl_context {
    // whether fuse moe combine
    cl_uint fuse_moe_combine;

+    // whether to fold the MoE bias adds into swiglu_oai
+    cl_uint fuse_moe_bias_glu;
+
+    // whether to fold the MoE down-projection bias add into the combine
+    cl_uint fuse_moe_bias_combine;
+
    bool adreno_has_large_buffer;
    bool adreno_use_large_buffer;
    bool adreno_use_bin_kernels;
    get_adreno_bin_kernel_func_t get_adreno_bin_kernel_func = nullptr;
    ggml_cl_compiler_version adreno_cl_compiler_version;
+    // The q6_K flat mul_mat codegen workarounds are needed by old E031 compilers only.
+    bool q6_k_flat_old_compiler;

    std::string kernel_compile_opts;  // cached for lazy-compiled kernels.

@@ -656,6 +664,7 @@ struct ggml_backend_opencl_context {

    cl_program program_add;
    cl_program program_add_id;
+    cl_program program_moe_add_id_glu;
    cl_program program_clamp;
    cl_program program_cvt;
    cl_program program_diag_mask_inf;
@@ -721,6 +730,7 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_div, kernel_div_row, kernel_div_f16, kernel_div_row_f16;
    cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
    cl_kernel kernel_add_id;
+    cl_kernel kernel_add_id_add_id_swiglu_oai;
    cl_kernel kernel_scale_f32, kernel_scale_f32_4;
    cl_kernel kernel_sqr_cont_f32, kernel_sqr_cont_f32_4, kernel_sqr_cont_f16, kernel_sqr_cont_f16_4;
    cl_kernel kernel_sqrt_cont_f32, kernel_sqrt_cont_f32_4, kernel_sqrt_cont_f16, kernel_sqrt_cont_f16_4;
@@ -866,6 +876,9 @@ struct ggml_backend_opencl_context {
    // [size_idx][kda][tgpp] where size_idx: 0=S_V=16, 1=32, 2=64, 3=128; kda: 0 or 1.
    // tgpp 0 = TG variant (COLS_PER_LANE_GROUP=1), tgpp 1 = prefill variant (COLS_PER_LANE_GROUP=4).
    cl_kernel kernel_gated_delta_net_f32[4][2][2] = {};
+    cl_kernel kernel_ssm_scan_f32_mamba2_d128 = nullptr;
+    cl_kernel kernel_ssm_scan_f32_mamba2_d256 = nullptr;
+
    cl_kernel kernel_timestep_embedding;
    cl_kernel kernel_gemv_moe_q4_0_f32_ns, kernel_gemm_moe_q4_0_f32_ns, kernel_gemm_moe_q4_0_f32_ns_bin;
    cl_kernel kernel_gemm_moe_q8_0_f32_ns;
@@ -892,7 +905,9 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_gemm_moe_q4_0_q8_1_dp4a = nullptr;    // dp4a (int8) q4_0 MoE prefill GEMM
    cl_kernel kernel_moe_reorder_b;
    cl_kernel kernel_moe_histogram, kernel_moe_scan, kernel_moe_fill, kernel_moe_scatter;
+    cl_kernel kernel_moe_scatter_stable = nullptr;   // deterministic slot assignment
    cl_kernel kernel_moe_combine_f32 = nullptr;   // fused router-weight mul + cross-expert sum
+    cl_kernel kernel_moe_combine_bias_f32 = nullptr;  // same, with the down-projection bias add folded in
    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
    cl_kernel kernel_mul_mv_id_q8_0_f32, kernel_mul_mv_id_q8_0_f32_flat;
    cl_kernel kernel_mul_mv_id_mxfp4_f32;
@@ -1340,6 +1355,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
        GGML_LOG_CONT(".");
    }

+    // moe_add_id_glu
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "moe_add_id_glu.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("moe_add_id_glu.cl");
+#endif
+        backend_ctx->program_moe_add_id_glu =
+            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_add_id_add_id_swiglu_oai =
+            clCreateKernel(backend_ctx->program_moe_add_id_glu, "kernel_add_id_add_id_swiglu_oai", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
    // tri
    {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1927,8 +1959,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
 #else
        const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl");
 #endif
+        // The codegen workarounds in this kernel are a measured 13-20% loss on
+        // compilers that do not need them, so only the affected ones build them;
+        // everyone else gets the original source.
+        const std::string q6k_opts = backend_ctx->q6_k_flat_old_compiler
+            ? compile_opts + " -DADRENO_OLD_COMPILER=1"
+            : compile_opts;
        cl_program prog =
-            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+            build_program_from_source(backend_ctx, kernel_src.c_str(), q6k_opts);

        CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err));
        CL_CHECK(clReleaseProgram(prog));
@@ -3154,6 +3192,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
        GGML_LOG_CONT(".");
    }

+    // ssm_scan (Mamba-2 fused per-token recurrent step; d_state in {128, 256})
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "ssm_scan.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("ssm_scan.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_ssm_scan_f32_mamba2_d128 = clCreateKernel(prog, "kernel_ssm_scan_f32_mamba2_d128", &err), err));
+        CL_CHECK((backend_ctx->kernel_ssm_scan_f32_mamba2_d256 = clCreateKernel(prog, "kernel_ssm_scan_f32_mamba2_d256", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
    // gated_delta_net: one kernel per (S_V, KDA, tgpp) triple.
    {
    #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -3246,6 +3302,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
            backend_ctx, kernel_src.c_str(), compile_opts);
        CL_CHECK((backend_ctx->kernel_moe_combine_f32 =
                    clCreateKernel(prog, "kernel_moe_combine_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_moe_combine_bias_f32 =
+                    clCreateKernel(prog, "kernel_moe_combine_bias_f32", &err), err));
        CL_CHECK(clReleaseProgram(prog));
        GGML_LOG_CONT(".");
    }
@@ -4442,6 +4500,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
        CL_CHECK((backend_ctx->kernel_moe_scan = clCreateKernel(prog, "kernel_moe_scan", &err), err));
        CL_CHECK((backend_ctx->kernel_moe_fill = clCreateKernel(prog, "kernel_moe_fill", &err), err));
        CL_CHECK((backend_ctx->kernel_moe_scatter = clCreateKernel(prog, "kernel_moe_scatter", &err), err));
+        CL_CHECK((backend_ctx->kernel_moe_scatter_stable = clCreateKernel(prog, "kernel_moe_scatter_stable", &err), err));
        CL_CHECK(clReleaseProgram(prog));
        GGML_LOG_CONT(".");
    }
@@ -5894,6 +5953,16 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
        (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
        (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);

+    // The q6_K flat mul_mat miscompile is a defect of the older E031 compilers, not a
+    // property of any GPU generation: it reproduces on E031.38 (Adreno 642L) and E031.41
+    // (Adreno 740) and is fixed by E031.45 (Adreno 619). Gate on the compiler so parts
+    // that do not need the workarounds do not pay for them. The explicit type check is
+    // required: newer_than_or_same() is false for every non-E031 compiler, so negating it
+    // alone would enable the workarounds on E17/DX.
+    backend_ctx->q6_k_flat_old_compiler =
+        backend_ctx->adreno_cl_compiler_version.type == E031 &&
+        !backend_ctx->adreno_cl_compiler_version.newer_than_or_same(E031, 45, 0, 0);
+
    size_t ext_str_size;
    clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
    char *ext_buffer = (char *)alloca(ext_str_size + 1);
@@ -5971,6 +6040,12 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
    backend_ctx->adreno_moe_ragged_skip_gran = (ragged_gran_env != NULL) ? atoi(ragged_gran_env) : 8;

    // whether fuse moe combine
+    static const char * fuse_moe_bias_glu_env = getenv("GGML_OPENCL_FUSE_MOE_BIAS_GLU");
+    backend_ctx->fuse_moe_bias_glu = fuse_moe_bias_glu_env == NULL ? 1 : (atoi(fuse_moe_bias_glu_env) != 0);
+
+    static const char * fuse_moe_bias_combine_env = getenv("GGML_OPENCL_FUSE_MOE_BIAS_COMBINE");
+    backend_ctx->fuse_moe_bias_combine = fuse_moe_bias_combine_env == NULL ? 1 : (atoi(fuse_moe_bias_combine_env) != 0);
+
    static const char * fuse_moe_combine_env = getenv("GGML_OPENCL_FUSE_MOE_COMBINE");
    backend_ctx->fuse_moe_combine = fuse_moe_combine_env == NULL ? 1 : (atoi(fuse_moe_combine_env) != 0);

@@ -6839,6 +6914,300 @@ static bool ggml_opencl_can_fuse_moe_combine(const struct ggml_cgraph * cgraph,
    return true;
 }

+// Detect the gpt-oss MoE bias+activation epilogue on the PREFILL path:
+// {MUL_MAT_ID(gate), ADD_ID(gate_bias), MUL_MAT_ID(up), ADD_ID(up_bias), GLU(swiglu_oai)}.
+// The two matmuls still run as their own dispatches (the prefill GEMM is the vendor's);
+// what collapses is the epilogue — both add_id passes are in-place read-modify-writes of a
+// tensor the GLU immediately reads again, so they are three full passes over the same
+// [n_ff, n_expert_used, n_tokens] f32 tensor where one suffices.
+//
+// The decode counterpart is handled by the mxfp4 fused GEMV arm in ggml_opencl_can_fuse,
+// which folds the matmul too; this one deliberately fires only when that cannot (ne[2] > 1).
+static bool ggml_opencl_can_fuse_moe_bias_glu(const struct ggml_cgraph * cgraph, int node_idx) {
+    if (node_idx + 4 >= cgraph->n_nodes) {
+        return false;
+    }
+
+    const enum ggml_op mg_ops[] = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU };
+    const int          mg_out[] = { node_idx + 4 };
+    if (!ggml_can_fuse_subgraph(cgraph, node_idx, 5, mg_ops, mg_out, 1)) {
+        return false;
+    }
+
+    const ggml_tensor * gmm = cgraph->nodes[node_idx];
+    const ggml_tensor * gad = cgraph->nodes[node_idx+1];
+    const ggml_tensor * umm = cgraph->nodes[node_idx+2];
+    const ggml_tensor * uad = cgraph->nodes[node_idx+3];
+    const ggml_tensor * glu = cgraph->nodes[node_idx+4];
+
+    if (ggml_get_glu_op(glu) != GGML_GLU_OP_SWIGLU_OAI) {
+        return false;
+    }
+    // Prefill only — at one token the mxfp4 arm above folds the matmul as well.
+    if (gmm->src[1]->ne[2] == 1) {
+        return false;
+    }
+    // Wiring: both matmuls share the activation and the expert selection, each add_id
+    // biases its own matmul, and the GLU consumes the two biased results as separate
+    // operands (so the same-buffer ne00_off/ne10_off split path is not in play).
+    if (gad->src[0] != gmm || uad->src[0] != umm ||
+        glu->src[0] != gad || glu->src[1] != uad ||
+        umm->src[1] != gmm->src[1] || umm->src[2] != gmm->src[2]) {
+        return false;
+    }
+    // A swapped GLU would exchange the gate/up roles the fused kernel hard-codes.
+    if (ggml_get_op_params_i32(glu, 1)) {
+        return false;
+    }
+    if (gad->type != GGML_TYPE_F32 || uad->type != GGML_TYPE_F32 || glu->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!gad->src[1] || gad->src[1]->type != GGML_TYPE_F32 ||
+        !uad->src[1] || uad->src[1]->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!gad->src[2] || gad->src[2]->type != GGML_TYPE_I32 || uad->src[2] != gad->src[2]) {
+        return false;
+    }
+    // Full width on both operands: the kernel writes one output element per input pair.
+    if (!ggml_are_same_shape(gad, uad) || glu->ne[0] != gad->ne[0] ||
+        glu->ne[1] != gad->ne[1] || glu->ne[2] != gad->ne[2] || glu->ne[3] != gad->ne[3]) {
+        return false;
+    }
+    if (gad->ne[3] != 1) {
+        return false;
+    }
+    // The destination is addressed by (expert slot, token) rather than the GLU's flat row
+    // walk; those agree only for a contiguous destination.
+    if (!ggml_is_contiguous(glu) || !ggml_is_contiguous(gmm) || !ggml_is_contiguous(umm)) {
+        return false;
+    }
+    return true;
+}
+
+static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+
+// Runs the gate and up matmuls unchanged, then one kernel in place of
+// add_id(gate) + add_id(up) + swiglu_oai. See ggml_opencl_can_fuse_moe_bias_glu.
+static void ggml_cl_moe_bias_glu_fused(ggml_backend_t backend, ggml_tensor * gate_mm, const ggml_tensor * gate_add,
+                                       ggml_tensor * up_mm, const ggml_tensor * up_add, const ggml_tensor * glu) {
+    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_cl_mul_mat_id(backend, gate_mm->src[0], gate_mm->src[1], gate_mm);
+    ggml_cl_mul_mat_id(backend, up_mm->src[0], up_mm->src[1], up_mm);
+
+    const ggml_tensor * gbias = gate_add->src[1];
+    const ggml_tensor * ubias = up_add->src[1];
+    const ggml_tensor * ids   = gate_add->src[2];
+
+    ggml_tensor_extra_cl * eg  = (ggml_tensor_extra_cl *)gate_mm->extra;
+    ggml_tensor_extra_cl * egb = (ggml_tensor_extra_cl *)gbias->extra;
+    ggml_tensor_extra_cl * eu  = (ggml_tensor_extra_cl *)up_mm->extra;
+    ggml_tensor_extra_cl * eub = (ggml_tensor_extra_cl *)ubias->extra;
+    ggml_tensor_extra_cl * ei  = (ggml_tensor_extra_cl *)ids->extra;
+    ggml_tensor_extra_cl * ed  = (ggml_tensor_extra_cl *)glu->extra;
+
+    cl_ulong off_g  = eg->offset  + gate_mm->view_offs;
+    cl_ulong off_gb = egb->offset + gbias->view_offs;
+    cl_ulong off_u  = eu->offset  + up_mm->view_offs;
+    cl_ulong off_ub = eub->offset + ubias->view_offs;
+    cl_ulong off_i  = ei->offset  + ids->view_offs;
+    cl_ulong off_d  = ed->offset  + glu->view_offs;
+
+    const cl_ulong nb01_g = gate_mm->nb[1];
+    const cl_ulong nb02_g = gate_mm->nb[2];
+    const cl_ulong nb01_u = up_mm->nb[1];
+    const cl_ulong nb02_u = up_mm->nb[2];
+    const cl_ulong nb11_g = gbias->nb[1];
+    const cl_ulong nb11_u = ubias->nb[1];
+    const cl_ulong nb21   = ids->nb[1];
+    const cl_ulong nbd1   = glu->nb[1];
+    const cl_ulong nbd2   = glu->nb[2];
+
+    const int   ne0   = (int)glu->ne[0];
+    const float alpha = ggml_get_op_params_f32(glu, 2);
+    const float limit = ggml_get_op_params_f32(glu, 3);
+
+    cl_kernel kernel = backend_ctx->kernel_add_id_add_id_swiglu_oai;
+
+    int i = 0;
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &eg->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &egb->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_gb));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &eu->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &eub->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_ub));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &ei->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_i));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &ed->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_d));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb01_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb02_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb01_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb02_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb11_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb11_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nbd1));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nbd2));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(float),    &limit));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(float),    &alpha));
+
+    const int nth = MIN(ne0, (int) backend_ctx->get_kernel_workgroup_size(kernel));
+    size_t global_work_size[] = { (size_t)glu->ne[1]*nth, (size_t)glu->ne[2], 1 };
+    size_t local_work_size[]  = { (size_t)nth, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, (ggml_tensor *)glu);
+}
+
+// Fusion B: the MoE down-projection bias add feeding the combine.
+//
+// The graph runs ADD_ID(down_bias) and then immediately the combine subgraph
+// {MUL(router weights), k VIEWs, k-1 ADDs}, and the ADD_ID's only consumer is that
+// MUL. Since the ADD_ID is an in-place read-modify-write of a tensor the combine
+// reads once more, the bias can be added inside the combine instead, dropping a
+// full pass over [n_embd, k, n_tokens].
+//
+// Shape checks for the combine tail are delegated to ggml_opencl_can_fuse_moe_combine
+// (which also owns the n_nodes >= 32 bail and the experts/dst aliasing bail); what is
+// added here is the ADD_ID wiring plus a subgraph check over the WHOLE run, so that
+// the intermediate bias result is confirmed not to escape.
+static bool ggml_opencl_can_fuse_moe_bias_combine(const struct ggml_cgraph * cgraph, int node_idx,
+                                                  const ggml_tensor ** out_final_add) {
+    if (node_idx + 1 >= cgraph->n_nodes) {
+        return false;
+    }
+    const ggml_tensor * add = cgraph->nodes[node_idx];
+    if (add->op != GGML_OP_ADD_ID) {
+        return false;
+    }
+    const ggml_tensor * mul = cgraph->nodes[node_idx+1];
+    if (mul->op != GGML_OP_MUL || mul->src[0] != add) {
+        return false;
+    }
+
+    const ggml_tensor * final_add = NULL;
+    if (!ggml_opencl_can_fuse_moe_combine(cgraph, node_idx+1, &final_add)) {
+        return false;
+    }
+
+    const ggml_tensor * raw  = add->src[0];
+    const ggml_tensor * bias = add->src[1];
+    const ggml_tensor * ids  = add->src[2];
+    if (!raw || !bias || !ids) {
+        return false;
+    }
+    if (raw->type != GGML_TYPE_F32 || bias->type != GGML_TYPE_F32 ||
+        ids->type != GGML_TYPE_I32 || add->type != GGML_TYPE_F32) {
+        return false;
+    }
+    // The combine reads the raw matmul output with the strides it computed from the
+    // add_id result, so the two must have the same layout.
+    if (!ggml_are_same_shape(raw, add) || !ggml_is_contiguous(raw)) {
+        return false;
+    }
+    if (raw->nb[1] != add->nb[1] || raw->nb[2] != add->nb[2]) {
+        return false;
+    }
+    // ids is indexed as [expert slot, token]; the combine walks the same two axes.
+    if (ids->ne[0] < add->ne[1] || ids->ne[1] < add->ne[2]) {
+        return false;
+    }
+
+    // Whole-run escape check: ADD_ID + MUL + k VIEWs + (k-1) ADDs, only the last node escapes.
+    const int k       = (int)add->ne[1];
+    const int n_nodes = 2 + k + (k - 1);
+    if (n_nodes >= 32 || node_idx + n_nodes > cgraph->n_nodes) {
+        return false;
+    }
+    enum ggml_op ops[32];
+    int n = 0;
+    ops[n++] = GGML_OP_ADD_ID;
+    ops[n++] = GGML_OP_MUL;
+    for (int j = 0; j < k;     ++j) ops[n++] = GGML_OP_VIEW;
+    for (int j = 0; j < k - 1; ++j) ops[n++] = GGML_OP_ADD;
+    const int outs[] = { node_idx + n_nodes - 1 };
+    if (!ggml_can_fuse_subgraph(cgraph, node_idx, n_nodes, ops, outs, 1)) {
+        return false;
+    }
+
+    *out_final_add = final_add;
+    return true;
+}
+
+
+// Fusion B dispatch: the combine, reading the RAW matmul output and adding the
+// per-expert bias row inline. See ggml_opencl_can_fuse_moe_bias_combine.
+static void ggml_cl_moe_bias_combine_fused(ggml_backend_t backend, const ggml_tensor * add,
+                                           const ggml_tensor * mul, const ggml_tensor * dst) {
+    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const ggml_tensor * experts = add->src[0];   // raw matmul output, bias not yet applied
+    const ggml_tensor * bias    = add->src[1];
+    const ggml_tensor * ids     = add->src[2];
+    const ggml_tensor * weights = mul->src[1];
+
+    ggml_tensor_extra_cl * ee = (ggml_tensor_extra_cl *)experts->extra;
+    ggml_tensor_extra_cl * eb = (ggml_tensor_extra_cl *)bias->extra;
+    ggml_tensor_extra_cl * ei = (ggml_tensor_extra_cl *)ids->extra;
+    ggml_tensor_extra_cl * ew = (ggml_tensor_extra_cl *)weights->extra;
+    ggml_tensor_extra_cl * ed = (ggml_tensor_extra_cl *)dst->extra;
+    cl_ulong off_e = ee->offset + experts->view_offs;
+    cl_ulong off_b = eb->offset + bias->view_offs;
+    cl_ulong off_i = ei->offset + ids->view_offs;
+    cl_ulong off_w = ew->offset + weights->view_offs;
+    cl_ulong off_d = ed->offset + dst->view_offs;
+
+    const int n_embd4 = (int)(experts->ne[0] / 4);
+    const int k       = (int)experts->ne[1];
+    const int nt      = (int)experts->ne[2];
+    const cl_uint e1 = (cl_uint)(experts->nb[1] / sizeof(float));
+    const cl_uint e2 = (cl_uint)(experts->nb[2] / sizeof(float));
+    const cl_uint w1 = (cl_uint)(weights->nb[1] / sizeof(float));
+    const cl_uint w2 = (cl_uint)(weights->nb[2] / sizeof(float));
+    const cl_uint d1 = (cl_uint)(dst->nb[1] / sizeof(float));
+    const cl_ulong nb_b1 = bias->nb[1];
+    const cl_ulong nb_i1 = ids->nb[1];
+
+    const size_t w_bytes = ggml_nbytes(weights);
+    backend_ctx->prealloc_moe_combine_w.allocate(backend_ctx->context, w_bytes);
+    CL_CHECK(clEnqueueCopyBuffer(backend_ctx->queue, ew->data_device, backend_ctx->prealloc_moe_combine_w.buffer,
+                                 off_w, 0, w_bytes, 0, NULL, NULL));
+    cl_mem   w_dev = backend_ctx->prealloc_moe_combine_w.buffer;
+    cl_ulong w_off = 0;
+
+    cl_kernel kernel = backend_ctx->kernel_moe_combine_bias_f32;
+    int a = 0;
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &ee->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_e));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &w_dev));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &w_off));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &eb->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_b));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &ei->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_i));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &ed->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_d));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(int),      &n_embd4));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(int),      &k));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(int),      &nt));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &e1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &e2));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &w1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &w2));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &d1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &nb_b1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &nb_i1));
+
+    size_t lws[2] = { 64, 1 };
+    size_t gws[2] = { (size_t)(((n_embd4 + 63) / 64) * 64), (size_t)nt };
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, gws, lws, (ggml_tensor *)dst);
+}
+
+
 static void ggml_cl_moe_combine_fused(ggml_backend_t backend, const ggml_tensor * mul, const ggml_tensor * dst) {
    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *)backend->context;
    const ggml_tensor * experts = mul->src[0];
@@ -6993,6 +7362,31 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
        }
        // Fuse the MoE combine: router-weight mul + cross-expert add chain ->
        // one weighted-sum-across-experts kernel.
+        // Fold the gpt-oss MoE bias epilogue: add_id(gate_bias) + add_id(up_bias) +
+        // glu(swiglu_oai) -> one kernel, leaving the two matmuls as their own dispatches.
+        // Both add_ids are in-place passes over a tensor the GLU reads again, so this
+        // drops two full read+write passes per layer. Opt out GGML_OPENCL_FUSE_MOE_BIAS_GLU=0.
+        if (backend_ctx->fuse_moe_bias_glu && !backend_ctx->disable_fusion &&
+            ggml_opencl_can_fuse_moe_bias_glu(cgraph, i)) {
+            ggml_cl_moe_bias_glu_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2],
+                                       cgraph->nodes[i+3], cgraph->nodes[i+4]);
+            i += 4;
+            continue;
+        }
+
+        // Fold the MoE down-projection bias into the combine: add_id(down_bias) + the whole
+        // combine subgraph -> one kernel. Checked before the plain combine arm so the longer
+        // pattern wins. Opt out GGML_OPENCL_FUSE_MOE_BIAS_COMBINE=0.
+        if (backend_ctx->fuse_moe_bias_combine && backend_ctx->fuse_moe_combine &&
+            !backend_ctx->disable_fusion) {
+            const ggml_tensor * bias_combine_out = nullptr;
+            if (ggml_opencl_can_fuse_moe_bias_combine(cgraph, i, &bias_combine_out)) {
+                ggml_cl_moe_bias_combine_fused(backend, node, cgraph->nodes[i+1], bias_combine_out);
+                i += 2 * (int)node->ne[1];   // ADD_ID + MUL + k VIEWs + (k-1) ADDs
+                continue;
+            }
+        }
+
        if (backend_ctx->fuse_moe_combine && !backend_ctx->disable_fusion) {
            const ggml_tensor * combine_out = nullptr;
            if (ggml_opencl_can_fuse_moe_combine(cgraph, i, &combine_out)) {
@@ -7301,6 +7695,23 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                   (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
        case GGML_OP_SSM_CONV:
            return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
+        case GGML_OP_SSM_SCAN: {
+            // Mamba-2 fused per-token scan. Requires src3->ne[0] == 1 (scalar
+            // A per head); d_state in {128, 256}; all sources f32. Falls back
+            // to CPU otherwise (incl. Mamba-1 element-wise A).
+            for (int i = 0; i < 6; ++i) {
+                if (op->src[i]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+            }
+            if (op->type != GGML_TYPE_F32) {
+                return false;
+            }
+            const int K = ggml_get_op_params_i32(op, 0);
+            const int d_state = (int) op->src[0]->ne[0];
+            const bool is_mamba2 = (op->src[3]->ne[0] == 1);
+            return is_mamba2 && (d_state == 128 || d_state == 256) && (K == 1);
+        }
        case GGML_OP_GATED_DELTA_NET:
            {
                // Match the Vulkan backend: only F32 -> F32, S_v in {16, 32, 64, 128}.
@@ -7335,6 +7746,19 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                       op->src[0]->type == GGML_TYPE_Q4_K  ||
                       op->src[0]->type == GGML_TYPE_Q5_K  ||
                       op->src[0]->type == GGML_TYPE_Q6_K) {
+                // The E031.41 compiler (usually with A7x) miscompiles the flat K-quant
+                // GEMV kernels (kernel_mul_mv_q*_K_f32_flat) and makes lm_head run much
+                // slower than it should. So, make it fallback to CPU to preserve performance
+                // for this compiler series.
+                static const char * a7x_lmhead_env = getenv("GGML_OPENCL_A7X_LMHEAD_CPU");
+                static const bool   a7x_lmhead_cpu = (a7x_lmhead_env == nullptr || a7x_lmhead_env[0] != '0');
+                if (a7x_lmhead_cpu &&
+                    backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                    (op->src[0]->type == GGML_TYPE_Q4_K || op->src[0]->type == GGML_TYPE_Q5_K ||
+                     op->src[0]->type == GGML_TYPE_Q6_K) &&
+                    op->src[0]->ne[1] >= 32768) {   // vocab-scale weight; no FFN/attn weight is this tall
+                    return false;
+                }
                return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
            } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
                return op->src[1]->type == GGML_TYPE_F32;
@@ -7453,6 +7877,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                     v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
            const bool is_f32_f16  = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
                                     v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
            const bool is_f32_q8_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q8_0 &&
                                     v->type == GGML_TYPE_Q8_0 && op->type == GGML_TYPE_F32 &&
                                     dk % 32 == 0 && dv % 32 == 0;
@@ -7460,6 +7885,21 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                     v->type == GGML_TYPE_Q4_0 && op->type == GGML_TYPE_F32 &&
                                     dk % 32 == 0 && dv % 32 == 0;

+            // A7X (Adreno 740, compiler E031.41) SIGSEGVs inside clBuildProgram
+            // building the flash_attn programs whose KV path is mixed-type or
+            // dequantized — f32_f16, q8_0, q4_0 (reproduced at DK=40 and DK=64; it
+            // is DK-independent). It is a driver crash, not codegen-wrong-output, so
+            // it cannot be caught in-process (fatal=false only handles clean compile
+            // errors). The uniform f16_f16 / f32_f32 programs compile fine on this
+            // compiler, so decline only the KV-convert variants; ggml then runs
+            // those (f16-KV / quant-KV) attention layers on the CPU backend.
+            // Negative compiler carve-out, same idiom as the Intel DK=512 decline
+            // below and the X1E driver-quirk guards.
+            if (backend_ctx && backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                (is_f32_f16 || is_f32_q8_0 || is_f32_q4_0)) {
+                return false;
+            }
+
            // Asymmetric KV: host-dequants both sides to F32, uses f32 kernel.
            auto is_kv_type_ok = [](ggml_type t) {
                return t == GGML_TYPE_F16 || t == GGML_TYPE_F32 ||
@@ -12257,6 +12697,103 @@ static void ggml_cl_mean(ggml_backend_t backend, const ggml_tensor * src0, const
    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }

+static void ggml_cl_ssm_scan(ggml_backend_t backend, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // s
+    const ggml_tensor * src1 = dst->src[1]; // x
+    const ggml_tensor * src2 = dst->src[2]; // dt
+    const ggml_tensor * src3 = dst->src[3]; // A
+    const ggml_tensor * src4 = dst->src[4]; // B
+    const ggml_tensor * src5 = dst->src[5]; // C
+    const ggml_tensor * src6 = dst->src[6]; // ids
+
+    GGML_ASSERT(src0 && src1 && src2 && src3 && src4 && src5 && src6 && dst);
+
+    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *) backend->context;
+
+    ggml_tensor_extra_cl * e0 = (ggml_tensor_extra_cl *) src0->extra;
+    ggml_tensor_extra_cl * e1 = (ggml_tensor_extra_cl *) src1->extra;
+    ggml_tensor_extra_cl * e2 = (ggml_tensor_extra_cl *) src2->extra;
+    ggml_tensor_extra_cl * e3 = (ggml_tensor_extra_cl *) src3->extra;
+    ggml_tensor_extra_cl * e4 = (ggml_tensor_extra_cl *) src4->extra;
+    ggml_tensor_extra_cl * e5 = (ggml_tensor_extra_cl *) src5->extra;
+    ggml_tensor_extra_cl * e6 = (ggml_tensor_extra_cl *) src6->extra;
+    ggml_tensor_extra_cl * ed = (ggml_tensor_extra_cl *) dst->extra;
+
+    cl_ulong o0 = e0->offset + src0->view_offs;
+    cl_ulong o1 = e1->offset + src1->view_offs;
+    cl_ulong o2 = e2->offset + src2->view_offs;
+    cl_ulong o3 = e3->offset + src3->view_offs;
+    cl_ulong o4 = e4->offset + src4->view_offs;
+    cl_ulong o5 = e5->offset + src5->view_offs;
+    cl_ulong o6 = e6->offset + src6->view_offs;
+    cl_ulong od = ed->offset + dst->view_offs;
+
+    const int d_state  = (int) src0->ne[0];
+    const int head_dim = (int) src0->ne[1];
+    const int n_head   = (int) src1->ne[1];
+    const int n_group  = (int) src4->ne[1];
+    const int n_tokens = (int) src1->ne[2];
+    const int n_seqs   = (int) src1->ne[3];
+
+    // Mirror CPU ref: s_off = ggml_nelements(src1) * sizeof(float)
+    const cl_ulong s_off_bytes = (cl_ulong) ggml_nelements(src1) * sizeof(float);
+
+    cl_kernel kernel = (d_state == 128)
+        ? backend_ctx->kernel_ssm_scan_f32_mamba2_d128
+        : backend_ctx->kernel_ssm_scan_f32_mamba2_d256;
+    GGML_ASSERT(kernel != nullptr);
+
+    cl_ulong s0_nb2 = src0->nb[2];
+    cl_ulong s0_nb3 = src0->nb[3];
+    cl_ulong x_nb2  = src1->nb[2];
+    cl_ulong x_nb3  = src1->nb[3];
+    cl_ulong dt_nb1 = src2->nb[1];
+    cl_ulong dt_nb2 = src2->nb[2];
+    cl_ulong A_nb1  = src3->nb[1];
+    cl_ulong B_nb2  = src4->nb[2];
+    cl_ulong B_nb3  = src4->nb[3];
+    cl_ulong C_nb2  = src5->nb[2];
+    cl_ulong C_nb3  = src5->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &e0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &o0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &e1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &o1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &e2->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &o2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &e3->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &o3));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_mem),   &e4->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &o4));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_mem),   &e5->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &o5));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_mem),   &e6->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &o6));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_mem),   &ed->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &od));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &s0_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &s0_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &x_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &x_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &dt_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &dt_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &A_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &B_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &B_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong), &C_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &C_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &s_off_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &head_dim));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(int),      &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_group));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(int),      &n_tokens));
+
+    size_t global_work_size[] = { (size_t)n_head * head_dim * 64, (size_t)n_seqs, 1 };
+    size_t local_work_size[]  = { 64, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_ssm_conv(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0);
    GGML_ASSERT(src0->extra);
@@ -12690,7 +13227,10 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);

-    const int nth = MIN(64, ne00);
+    int nth = 1;
+    while (nth < ne00 && nth < 64) {
+        nth *= 2;
+    }

    cl_kernel kernel = backend_ctx->kernel_norm;

@@ -20440,6 +20980,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne1));
            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r2));
            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r3));
+            // The optimizer-barrier arg exists only in the ADRENO_OLD_COMPILER build of
+            // this kernel; conformant compilers get the original 17-arg signature.
+            if (backend_ctx->q6_k_flat_old_compiler) {
+                cl_uchar q6k_mask = 0xFF;   // never 0xFE in prod; see the kernel note
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_uchar), &q6k_mask));
+            }
 #else
            kernel = backend_ctx->kernel_mul_mv_q6_K_f32;

@@ -20725,18 +21271,42 @@ static void moe_router_reoerder(ggml_backend_t backend, const ggml_tensor * src,
    size_t fill_local_size[] = {64, 1, 1};
    backend_ctx->enqueue_ndrange_kernel(kernel, 3, fill_global_size, fill_local_size, src);

-    // Scatter
-    kernel = backend_ctx->kernel_moe_scatter;
-    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
-    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
-    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
-    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
-    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &slot_counter_buf));
-    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne21));
-    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne20));
-    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne02));
+    // Scatter. The deterministic variant is the default: kernel_moe_scatter derives
+    // each token's slot from an atomic counter, so the packing inside an expert - and
+    // with it the output of the ragged prefill GEMM - changes from run to run. Set
+    // GGML_OPENCL_MOE_STABLE_SCATTER=0 to restore the atomic version.
+    static const bool stable_scatter = []{
+        const char * e = getenv("GGML_OPENCL_MOE_STABLE_SCATTER");
+        return !e || e[0] == '\0' || e[0] != '0';
+    }();

-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, histogram_global_size, histogram_local_size, src);
+    if (stable_scatter) {
+        kernel = backend_ctx->kernel_moe_scatter_stable;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne21));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne20));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne02));
+
+        // one workgroup (one wave) per expert; each ranks its own tokens
+        size_t scatter_global_size[] = {64, (size_t)ne02};
+        size_t scatter_local_size[]  = {64, 1};
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, scatter_global_size, scatter_local_size, src);
+    } else {
+        kernel = backend_ctx->kernel_moe_scatter;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &slot_counter_buf));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne21));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne20));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne02));
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, histogram_global_size, histogram_local_size, src);
+    }

    // [MOE_TILES] env-gated padding probe: read back total_tiles (= Sum_e
    // ceil(k_e/n_tile_size)) and compare to the ideal tile count for the real
@@ -23703,6 +24273,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    const int n_dims     = ((int *) dst->op_params)[1];
    const int mode       = ((int *) dst->op_params)[2];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];

    float freq_base;
    float freq_scale;
@@ -23731,6 +24302,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const

    if (is_vision) {
        GGML_ASSERT(n_dims == ne00/2);
+        GGML_ASSERT(n_offs == 0); // offset not supported for vision, as the rotated pairs span the whole row
    }

    cl_kernel kernel;
@@ -23822,6 +24394,12 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
    if (is_mrope && !is_vision) {
        CL_CHECK(clSetKernelArg(kernel, 34, sizeof(int), &is_imrope));
    }
+    // norm and neox have n_offs after beta_slow, mrope has it after is_imrope
+    if (!is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int), &n_offs));
+    } else if (is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 35, sizeof(int), &n_offs));
+    }

    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
    size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -24743,6 +25321,14 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
            }
            func = ggml_cl_ssm_conv;
            break;
+        case GGML_OP_SSM_SCAN:
+            if (!any_on_device) {
+                return false;
+            }
+            // SSM_SCAN has 7 source tensors, so it cannot use the standard
+            // (src0, src1, dst) func signature. Dispatch directly and return.
+            ggml_cl_ssm_scan(backend, tensor);
+            return true;
        case GGML_OP_GATED_DELTA_NET:
            if (!any_on_device) {
                return false;
@@ -118,6 +118,17 @@ __kernel void flash_attn_f16(
    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];

    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+#if WG_SIZE > FA_SG
+        // WAR on l_k/l_v: a thread that finishes the compute below early — either
+        // it skipped it (my_query_row >= n_q, the continue) or its subgroup simply
+        // ran ahead — wraps around and reloads the tiles while another subgroup is
+        // still reading them. Any WG that is exactly one lockstep subgroup
+        // (WG_SIZE == FA_SG) cannot diverge and hides this; a WG spanning multiple
+        // subgroups (Intel sg=32, or BLOCK_M > 64 on Adreno) corrupts the result.
+        // All threads reach this each iteration (no-op on the first), so it does
+        // not diverge with the continue. Compiled out when WG == one subgroup.
+        barrier(CLK_LOCAL_MEM_FENCE);
+#endif
        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
            const int row = i / DK_VEC;
            const int col = i % DK_VEC;
--- a/Show More
+++ b/Show More