ggml : fix msvc+clang ggml_vld1q_u32 (#28284 )

Revert "ggml-cuda : restore prop.integrated on HIP builds (#24233 )" (#28604 )
This reverts commit c7d8722922.
2026-09-08 17:07:09 +02:00 · 2026-09-08 17:40:26 +03:00 · 2026-09-08 16:19:53 +02:00 · 2026-09-08 16:01:03 +03:00 · 2026-09-08 15:54:42 +03:00 · 2026-09-08 20:44:33 +08:00
225 changed files with 12195 additions and 4022 deletions
@@ -31,7 +31,7 @@
    ]
    && blas.meta.available,
  useCuda ? config.cudaSupport,
-  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
+  useMetalKit ? stdenv.hostPlatform.isAarch64 && stdenv.hostPlatform.isDarwin,
  # Increases the runtime closure size by ~700M
  useMpi ? false,
  useRocm ? config.rocmSupport,
@@ -92,7 +92,7 @@ let

  cudaBuildInputs = with cudaPackages; [
    cuda_cudart
-    cuda_cccl # <nv/target>
+    cccl # <nv/target>
    libcublas
  ];

@@ -166,7 +166,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
  # `xcrun` is used find the path of the Metal compiler, which is varible
  # and not on $PATH
  # see https://github.com/ggml-org/llama.cpp/pull/6118 for discussion
-  __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
+  __noChroot = effectiveStdenv.hostPlatform.isDarwin && useMetalKit && precompileMetalShaders;

  nativeBuildInputs =
    [
@@ -181,10 +181,10 @@ effectiveStdenv.mkDerivation (finalAttrs: {
      autoAddDriverRunpath
    ]
    ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static ]
-    ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
+    ++ optionals (effectiveStdenv.hostPlatform.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];

  buildInputs =
-    optionals effectiveStdenv.isDarwin darwinBuildInputs
+    optionals effectiveStdenv.hostPlatform.isDarwin darwinBuildInputs
    ++ optionals useCuda cudaBuildInputs
    ++ optionals useMpi [ mpi ]
    ++ optionals useRocm rocmBuildInputs
@@ -245,7 +245,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {

    # Configurations that are known to result in build failures. Can be
    # overridden by importing Nixpkgs with `allowBroken = true`.
-    broken = (useMetalKit && !effectiveStdenv.isDarwin);
+    broken = (useMetalKit && !effectiveStdenv.hostPlatform.isDarwin);

    description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
    homepage = "https://github.com/ggml-org/llama.cpp/";
@@ -1,4 +1,4 @@
-blank_issues_enabled: true
+blank_issues_enabled: false
 contact_links:
  - name: Got an idea?
    url: https://github.com/ggml-org/llama.cpp/discussions/categories/ideas
@@ -50,8 +50,16 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: apple-arm64
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: apple-arm64
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Build
        id: cmake_build
@@ -67,6 +75,18 @@ jobs:
            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: apple-arm64
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Check for leaks
        run: |
          cmd=(./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1)
@@ -80,16 +100,6 @@ jobs:
          cd build
          ctest -L main -E "test-llama-archs" --verbose --timeout 900

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: apple-arm64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
  macos-latest-x64:
    runs-on: macos-15-intel

@@ -102,8 +112,16 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: apple-x64
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: apple-x64
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Build
        id: cmake_build
@@ -120,22 +138,24 @@ jobs:
            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: apple-x64
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Test
        id: cmake_test
        run: |
          cd build
          ctest -L main --verbose --timeout 900

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: apple-x64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
  macos-latest-ios-xcode:
    runs-on: macos-latest

@@ -65,8 +65,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: cpu-${{ matrix.os }}
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false

      - name: Build Dependencies
        id: build_depends
@@ -91,6 +90,15 @@ jobs:
          python3 -m pip install --upgrade pip setuptools
          pip3 install ./gguf-py

+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: cpu-${{ matrix.os }}
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
      - name: Build
        id: cmake_build
        run: |
@@ -100,6 +108,18 @@ jobs:
            -DGGML_RPC=ON
          time cmake --build build --config Release -j $(nproc)

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: cpu-${{ matrix.os }}
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Test
        id: cmake_test
        run: |
@@ -117,18 +137,6 @@ jobs:
          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

-      # note: real deletion only on push to master (same condition as the ccache save),
-      #       dry-run otherwise (the token is read-only on PRs from forks)
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: cpu-${{ matrix.os }}
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
  windows:
    name: windows / ${{ matrix.build }}
    runs-on: windows-2025
@@ -61,7 +61,7 @@ jobs:
      - name: ccache-buckets-restore
        uses: ./.github/actions/ccache-buckets
        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
        with:
          key: cuda-ubuntu-24.04-cuda
          folder: llama.cpp
@@ -116,7 +116,7 @@ jobs:
      - name: ccache-buckets-restore
        uses: ./.github/actions/ccache-buckets
        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
        with:
          key: cuda-ubuntu-22.04-hip
          folder: llama.cpp
@@ -167,7 +167,7 @@ jobs:
      - name: ccache-buckets-restore
        uses: ./.github/actions/ccache-buckets
        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
        with:
          key: cuda-ubuntu-22.04-musa
          folder: llama.cpp
@@ -78,8 +78,16 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: sycl-ubuntu-24-${{ matrix.build }}
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: sycl-ubuntu-24-${{ matrix.build }}
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Build
        id: cmake_build
@@ -96,15 +104,17 @@ jobs:
            -DGGML_SYCL_F16=${{ matrix.fp16 }}
          time cmake --build build --config Release -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
        env:
-          GH_TOKEN: ${{ github.token }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
        with:
          key: sycl-ubuntu-24-${{ matrix.build }}
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true

  windows-latest-sycl:
    runs-on: windows-2022
@@ -57,8 +57,16 @@ jobs:
        with:
          key: vulkan-ubuntu-24.04-arm
          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: vulkan-ubuntu-24.04-arm
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Configure
        id: cmake_configure
@@ -73,15 +81,17 @@ jobs:
        run: |
          time cmake --build build -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
        env:
-          GH_TOKEN: ${{ github.token }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
        with:
          key: vulkan-ubuntu-24.04-arm
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true

  ubuntu-llvmpipe:
    runs-on: ubuntu-24.04
@@ -115,8 +125,16 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: vulkan-ubuntu-24.04-llvmpipe
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: vulkan-ubuntu-24.04-llvmpipe
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Build
        id: cmake_build
@@ -127,6 +145,18 @@ jobs:
            -DGGML_VULKAN=ON
          cmake --build build --config Release -j $(nproc)

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: vulkan-ubuntu-24.04-llvmpipe
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Test
        id: cmake_test
        run: |
@@ -138,16 +168,6 @@ jobs:
          # test-backend-ops is too slow on llvmpipe, skip it
          ctest -L main -E test-backend-ops --verbose --timeout 900

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: vulkan-ubuntu-24.04-llvmpipe
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
  windows:
    runs-on: windows-2025

@@ -57,8 +57,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: webgpu-ubuntu-24.04-arm-wasm
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false

      - name: Install Emscripten
        run: |
@@ -76,6 +75,15 @@ jobs:
            "https://github.com/google/dawn/releases/download/${DAWN_TAG}/${EMDAWN_PKG}"
          unzip emdawn.zip

+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: webgpu-ubuntu-24.04-arm-wasm
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
      - name: Build WASM WebGPU
        run: |
          source emsdk/emsdk_env.sh
@@ -89,12 +97,14 @@ jobs:

          time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
        env:
-          GH_TOKEN: ${{ github.token }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
        with:
          key: webgpu-ubuntu-24.04-arm-wasm
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
@@ -72,8 +72,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: webgpu-macos-latest
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false

      - name: Dawn Dependency
        id: dawn-depends
@@ -88,6 +87,15 @@ jobs:
          mkdir dawn
          tar -xvf artifact.tar.gz -C dawn --strip-components=1

+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: webgpu-macos-latest
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
      - name: Build
        id: cmake_build
        run: |
@@ -95,22 +103,24 @@ jobs:
          cmake -B build -G "Ninja" -DCMAKE_BUILD_TYPE=Release -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: webgpu-macos-latest
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Test
        id: cmake_test
        run: |
          cd build
          ctest -L main --verbose --timeout 900

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: webgpu-macos-latest
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
  ubuntu:
    runs-on: ubuntu-24.04

@@ -123,8 +133,7 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: webgpu-ubuntu-24.04
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false

      - name: Dependencies
        id: depends
@@ -148,6 +157,15 @@ jobs:
          mkdir dawn
          tar -xvf artifact.tar.gz -C dawn --strip-components=1

+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: webgpu-ubuntu-24.04
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
      - name: Build
        id: cmake_build
        run: |
@@ -156,6 +174,18 @@ jobs:
            -DGGML_WEBGPU=ON
          time cmake --build build --config Release -j $(nproc)

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: webgpu-ubuntu-24.04
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Test
        id: cmake_test
        run: |
@@ -163,13 +193,3 @@ 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
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: webgpu-ubuntu-24.04
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
@@ -49,14 +49,22 @@ jobs:
        id: depends
        run: |
          sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev python3
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev python3 python3-venv python3-pip jq

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: hip-quality-check-ubuntu-22.04
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: hip-quality-check-ubuntu-22.04
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Build with Werror
        id: cmake_build
@@ -85,12 +93,14 @@ jobs:
          make -j $(nproc) 2>&1 | tee metrics.log | grep -v 'Rpass-analysis=kernel-resource-usage\|remark:\|^$'
          python3 ../scripts/hip/gcn-cdna-vgpr-check.py metrics.log

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
        env:
-          GH_TOKEN: ${{ github.token }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
        with:
          key: hip-quality-check-ubuntu-22.04
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
@@ -19,6 +19,7 @@ env:

 permissions:
  contents: write
+  packages: write

 jobs:
  make-release:
@@ -113,6 +114,29 @@ jobs:
              data: await fs.readFileSync('./nightly-tag.txt')
            });

+      - name: Re-tag container images with release version
+        if: ${{ github.event.inputs.dry_run == 'false' && steps.desc.outputs.nightly_tag != '' }}
+        env:
+          GITHUB_REPOSITORY_OWNER: ${{ github.repository_owner }}
+        run: |
+          VERSION="${{ steps.checks.outputs.version }}"
+          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER,,}"
+          IMAGE_REPO="ghcr.io/${REPO_OWNER}/${{ github.event.repository.name }}"
+
+          echo "${{ secrets.GITHUB_TOKEN }}" | docker login ghcr.io -u "${{ github.actor }}" --password-stdin
+
+          VARIANTS=("" "-cuda" "-cuda13" "-vulkan" "-rocm" "-intel" "-musa" "-openvino")
+          TYPES=("full" "light" "server")
+          for type in "${TYPES[@]}"; do
+            for variant in "${VARIANTS[@]}"; do
+              src="${IMAGE_REPO}:${type}${variant}-${NIGHTLY_TAG}"
+              dst="${IMAGE_REPO}:${type}${variant}-${VERSION}"
+              echo "Tagging ${src} -> ${dst}"
+              docker buildx imagetools create --tag "${dst}" "${src}"
+            done
+          done
+
      - name: Dry run summary
        if: ${{ github.event.inputs.dry_run == 'true' }}
        run: |
@@ -31,7 +31,7 @@ jobs:
        uses: actions/setup-python@v6
        with:
          python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt ty==0.0.35
+          pip-install: -r requirements/requirements-all.txt ty==0.0.78
      # - name: Type-check with Pyright
      #   uses: jakebailey/pyright-action@v2
      #   with:
@@ -40,4 +40,4 @@ jobs:
      #     warnings: true
      - name: Type-check with ty
        run: |
-            ty check --output-format=github
+            ty check --exit-zero-on-warning --output-format=github
@@ -72,7 +72,7 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Tests (GPUx1, backend-sampling)
        id: server_integration_tests_backend_sampling
@@ -81,7 +81,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Tests (GPUx2)
        id: server_integration_tests_gpu2
@@ -90,7 +90,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_METAL_DEVICES=2
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Tests (GPUx2, backend-sampling)
        id: server_integration_tests_gpu2_backend_sampling
@@ -99,7 +99,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

  server-cuda:
    runs-on: "hf-jobs-t4-small:cuda13"
@@ -162,7 +162,7 @@ jobs:
        run: |
          cd tools/server/tests
          source venv/bin/activate
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Tests (GPUx1, backend-sampling)
        id: server_integration_tests_backend_sampling
@@ -171,7 +171,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Tests (GPUx2)
        id: server_integration_tests_gpu2
@@ -180,7 +180,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_CUDA_DEVICES=2
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Tests (GPUx2, backend-sampling)
        id: server_integration_tests_gpu2_backend_sampling
@@ -189,7 +189,7 @@ jobs:
          cd tools/server/tests
          source venv/bin/activate
          export GGML_CUDA_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

  server-kleidiai:
    runs-on: ah-ubuntu_22_04-c8g_8x
@@ -83,8 +83,16 @@ jobs:
        uses: ggml-org/ccache-action@v1.2.24
        with:
          key: server-ubuntu-24.04-arm
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: server-ubuntu-24.04-arm
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Build
        id: cmake_build
@@ -93,6 +101,18 @@ jobs:
            -DGGML_SCHED_NO_REALLOC=ON
          cmake --build build --config Release -j $(nproc) --target llama-server

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: server-ubuntu-24.04-arm
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Python setup
        id: setup_python
        uses: actions/setup-python@v6
@@ -128,16 +148,6 @@ jobs:
          export LLAMA_ARG_BACKEND_SAMPLING=1
          SLOW_TESTS=1 ./tests.sh

-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: server-ubuntu-24.04-arm
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
  windows:
    runs-on: windows-2025

@@ -17,8 +17,9 @@ jobs:
        uses: actions/setup-node@v6
        with:
          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          # cache: "npm"
+          # cache-dependency-path: "tools/ui/package-lock.json"
+          package-manager-cache: false

      - name: Install dependencies
        run: npm ci
@@ -33,8 +33,9 @@ jobs:
        uses: actions/setup-node@v6
        with:
          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          # cache: "npm"
+          # cache-dependency-path: "tools/ui/package-lock.json"
+          package-manager-cache: false

      - name: Install dependencies
        run: npm ci
@@ -57,8 +57,9 @@ jobs:
        uses: actions/setup-node@v6
        with:
          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          # cache: "npm"
+          # cache-dependency-path: "tools/ui/package-lock.json"
+          package-manager-cache: false

      - name: Download built UI artifacts
        uses: actions/download-artifact@v6
@@ -114,8 +115,9 @@ jobs:
        uses: actions/setup-node@v6
        with:
          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          # cache: "npm"
+          # cache-dependency-path: "tools/ui/package-lock.json"
+          package-manager-cache: false

      - name: Install dependencies
        id: setup
@@ -53,7 +53,10 @@ endif()

 set(TARGET llama-common)

+include(parsers/sources.cmake)
+
 add_library(${TARGET}
+    ${LLAMA_CHAT_PARSERS_SOURCES}
    arg.cpp
    arg.h
    base64.hpp
@@ -894,6 +894,12 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
    postprocess_cpu_params(params.speculative.draft.cpuparams,       &params.cpuparams);
    postprocess_cpu_params(params.speculative.draft.cpuparams_batch, &params.cpuparams_batch);

+    // default the mmproj device to the global device selection if not set explicitly with -mmdev
+    if (params.mmproj_use_gpu && params.mmproj_device == nullptr && !params.devices.empty()) {
+        params.mmproj_device = params.devices.front();
+        params.mmproj_use_gpu = params.mmproj_device != nullptr;
+    }
+
    if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
    }
@@ -2610,7 +2616,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    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"
+        "device to use for multimodal projector (none = don't offload, default: follows --device)\n"
        "use --list-devices to see a list of available devices",
        [](common_params & params, const std::string & value) {
            if (value == "none") {
@@ -3901,6 +3907,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            common_log_set_file(common_log_main(), value.c_str());
        }
    ).set_env("LLAMA_ARG_LOG_FILE"));
+    add_opt(common_arg(
+        {"--log-jsonl"},
+        {"--no-log-jsonl"},
+        "Log as JSONL (one JSON object per line) to stdout, this also disables colored logging (default: disabled)",
+        [](common_params &, bool value) {
+            common_log_set_jsonl(common_log_main(), value);
+        }
+    ).set_env("LLAMA_ARG_LOG_JSONL"));
    add_opt(common_arg(
        {"--log-prompts-dir"}, "PATH",
        "Log prompts to directory (auto-created if not present; only used for debugging, default: disabled)",
@@ -4221,7 +4235,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING"));
    add_opt(common_arg(
        {"--spec-draft-device", "-devd", "--device-draft"}, "<dev1,dev2,..>",
-        "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
+        "comma-separated list of devices to use for offloading the draft model (none = don't offload, default: follows --device)\n"
        "use --list-devices to see a list of available devices",
        [](common_params & params, const std::string & value) {
            params.speculative.draft.devices = parse_device_list(value);
@@ -117,6 +117,7 @@ caps caps_get(jinja::program & prog) {

    JJ_DEBUG("%s\n", ">>> Running capability check: typed content");

+    bool checks_for_string = false;
    static const std::string content_marker = "STRING_MARKER";

    // case: typed content support
@@ -136,6 +137,10 @@ caps caps_get(jinja::program & prog) {
        [&](context &, bool success, value & messages, value &, const std::string & rendered) {
            auto & content = messages->at(0)->at("content");
            caps_print_stats(content, "messages[0].content");
+            if (has_op(content, "test_is_string")) {
+                // checked if content is string
+                checks_for_string = true;
+            }
            bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
            if (used_as_array) {
                // accessed as an array
@@ -151,6 +156,33 @@ caps caps_get(jinja::program & prog) {
        }
    );

+    if (checks_for_string) {
+        caps_try_execute(
+            prog,
+            [&]() {
+                // messages
+                return json::array({
+                    {
+                        {"role", "user"},
+                        {"content", json::array({
+                        })}
+                    }
+                });
+            },
+            nullptr, // ctx_fn
+            nullptr, // tools_fn
+            [&](context &, bool success, value & messages, value &, const std::string &) {
+                auto & content = messages->at(0)->at("content");
+                caps_print_stats(content, "messages[0].content");
+                bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
+                if (used_as_array && success) {
+                    // accessed as an array
+                    result.supports_typed_content = true;
+                }
+            }
+        );
+    }
+
    JJ_DEBUG("%s\n", ">>> Running capability check: system prompt");

    // case: system prompt support
@@ -412,12 +412,18 @@ value test_expression::execute_impl(context & ctx) {
        throw std::runtime_error("Invalid test expression");
    }

-    auto it = builtins.find("test_is_" + test_id);
-    JJ_DEBUG("Test expression %s '%s' %s (using function 'test_is_%s')", operand->type().c_str(), test_id.c_str(), negate ? "(negate)" : "", test_id.c_str());
+    const std::string test_name = "test_is_" + test_id;
+    auto it = builtins.find(test_name);
+    JJ_DEBUG("Test expression %s '%s' %s (using function '%s')", operand->type().c_str(), test_id.c_str(), negate ? "(negate)" : "", test_name.c_str());
    if (it == builtins.end()) {
        throw std::runtime_error("Unknown test '" + test_id + "'");
    }

+    if (ctx.is_get_stats) {
+        value_t::stats_t::mark_used(input);
+        input->stats.ops.insert(test_name);
+    }
+
    auto res = it->second(args);

    if (negate) {
@@ -1,5 +1,6 @@
 #include "common.h"
 #include "log.h"
+#include "json.h"

 #include <chrono>
 #include <condition_variable>
@@ -66,6 +67,17 @@ static const char* g_col[] = {
    "",
 };

+static const char * level_str(enum ggml_log_level level) {
+    switch (level) {
+        case GGML_LOG_LEVEL_DEBUG: return "debug";
+        case GGML_LOG_LEVEL_INFO:  return "info";
+        case GGML_LOG_LEVEL_WARN:  return "warn";
+        case GGML_LOG_LEVEL_ERROR: return "error";
+        case GGML_LOG_LEVEL_CONT:  return "cont";
+        default:                   return "none";
+    }
+}
+
 struct common_log_entry {
    enum ggml_log_level level {GGML_LOG_LEVEL_INFO};

@@ -74,6 +86,7 @@ struct common_log_entry {
    int64_t timestamp { 0 };
    bool is_end       { false }; // signals the worker thread to stop
    bool prefix       { false };
+    bool jsonl        { false };

    common_log_entry(size_t size = 256) : msg(size) { }

@@ -88,11 +101,23 @@ struct common_log_entry {

            fcur = stdout;

-            if (level != GGML_LOG_LEVEL_NONE) {
+            if (level != GGML_LOG_LEVEL_NONE && !jsonl) {
                fcur = stderr;
            }
        }

+        if (jsonl) {
+            common_json obj = {
+                {"type",  "log"},
+                {"time",  timestamp},
+                {"level", level_str(level)},
+                {"msg",   msg.data()},
+            };
+            fprintf(fcur, "%s\n", obj.dump_safe().c_str());
+            fflush(fcur);
+            return;
+        }
+
        if (level != GGML_LOG_LEVEL_NONE && level != GGML_LOG_LEVEL_CONT && prefix) {
            if (timestamp) {
                // [M.s.ms.us]
@@ -131,6 +156,7 @@ struct common_log {
        file       = nullptr;
        prefix     = false;
        timestamps = false;
+        jsonl      = false;
        running    = false;
        t_start    = t_us();

@@ -158,6 +184,7 @@ private:

    bool prefix;
    bool timestamps;
+    bool jsonl;
    bool running;

    int64_t t_start;
@@ -246,6 +273,7 @@ public:
        entry.is_end    = false;
        entry.level     = level;
        entry.prefix    = prefix;
+        entry.jsonl     = jsonl;
        entry.timestamp = 0;
        if (timestamps) {
            entry.timestamp = t_us() - t_start;
@@ -360,6 +388,12 @@ public:

        this->timestamps = timestamps;
    }
+
+    void set_jsonl(bool jsonl) {
+        std::lock_guard<std::mutex> lock(mtx);
+
+        this->jsonl = jsonl;
+    }
 };

 //
@@ -433,6 +467,10 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
    log->set_timestamps(timestamps);
 }

+void common_log_set_jsonl(struct common_log * log, bool jsonl) {
+    log->set_jsonl(jsonl);
+}
+
 void common_log_flush(struct common_log * log) {
    log->pause();
    log->resume();
@@ -91,6 +91,7 @@ void common_log_set_file      (struct common_log * log, const char * file); // n
 void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
 void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
 void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
+void common_log_set_jsonl     (struct common_log * log, bool jsonl);        // print each log as a JSON object on one line, not thread-safe
 void common_log_flush         (struct common_log * log);                    // flush all pending log messages

 // helper macros for logging
@@ -0,0 +1,150 @@
+#include "parsers.h"
+
+// Cohere2 MoE (a.k.a. "North Code") parser.
+//
+// The assistant turn is fully marker-wrapped:
+//   <|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
+//     <|START_THINKING|>{reasoning}<|END_THINKING|>
+//     then EITHER content:    <|START_TEXT|>{content}<|END_TEXT|>
+//          OR     tool calls: <|START_ACTION|>[
+//                                 {"tool_call_id": "0", "tool_name": "f", "parameters": {...}}, ...
+//                             ]<|END_ACTION|>
+//   <|END_OF_TURN_TOKEN|>
+//
+// The generation prompt forces a leading <|START_THINKING|> (when reasoning is enabled, which is
+// the template default), so the model's output continues from *inside* the thinking block. The
+// parser literal therefore only covers the stable <|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|> prefix
+// and the reasoning rule consumes the <|START_THINKING|> ... <|END_THINKING|> markers itself,
+// regardless of whether they came from the generation prompt or the generated text.
+common_chat_params common_chat_params_init_cohere2moe(const common_chat_template &          tmpl,
+                                                              const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    const std::string TURN_START    = "<|START_OF_TURN_TOKEN|>";
+    const std::string TURN_END      = "<|END_OF_TURN_TOKEN|>";
+    const std::string CHATBOT       = "<|CHATBOT_TOKEN|>";
+    const std::string USER          = "<|USER_TOKEN|>";
+    const std::string SYSTEM        = "<|SYSTEM_TOKEN|>";
+    const std::string THINK_START   = "<|START_THINKING|>";
+    const std::string THINK_END     = "<|END_THINKING|>";
+    const std::string TEXT_START    = "<|START_TEXT|>";
+    const std::string TEXT_END      = "<|END_TEXT|>";
+    const std::string ACTION_START  = "<|START_ACTION|>";
+    const std::string ACTION_END    = "<|END_ACTION|>";
+    const std::string RESULT_START  = "<|START_TOOL_RESULT|>";
+    const std::string RESULT_END    = "<|END_TOOL_RESULT|>";
+
+    // Stable prefix of the generation prompt that precedes the (forced) <|START_THINKING|> marker.
+    const std::string GEN_PREFIX = TURN_START + CHATBOT;
+
+    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = {THINK_END};
+    data.preserved_tokens   = {
+        TURN_START, TURN_END, CHATBOT, USER, SYSTEM,
+        THINK_START, THINK_END,
+        TEXT_START, TEXT_END,
+        ACTION_START, ACTION_END,
+        RESULT_START, RESULT_END,
+    };
+
+    // Declare per-role message delimiters. Tool results are rendered with the
+    // system token followed by <|START_TOOL_RESULT|>, so the "tool" delimiter must be listed before
+    // the plain "system" one (it is a strict superset, and the role split tries delimiters in order).
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, GEN_PREFIX },
+        { COMMON_CHAT_ROLE_USER,      TURN_START + USER },
+        { COMMON_CHAT_ROLE_TOOL,      TURN_START + SYSTEM + RESULT_START },
+        { COMMON_CHAT_ROLE_SYSTEM,    TURN_START + SYSTEM },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PREFIX + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + TEXT_START + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal(GEN_PREFIX);
+        auto end               = p.end();
+
+        // The thinking block is always present (the generation prompt forces <|START_THINKING|>).
+        // When extracting reasoning, capture its body; otherwise keep the whole block (markers
+        // included) inline as content, matching reasoning_format=NONE conventions.
+        common_peg_parser reasoning = p.eps();
+        if (extract_reasoning) {
+            reasoning = p.optional(p.literal(THINK_START) +
+                                   p.reasoning(p.until_one_of({ THINK_END, TEXT_START, ACTION_START })) +
+                                   p.optional(p.literal(THINK_END)));
+        } else {
+            reasoning = p.optional(p.content(p.literal(THINK_START) +
+                                             p.until_one_of({ THINK_END, TEXT_START, ACTION_START }) +
+                                             p.optional(p.literal(THINK_END))));
+        }
+
+        auto text_content = has_response_format
+            ? p.literal(TEXT_START) +
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
+                p.optional(p.literal(TEXT_END))
+            : p.literal(TEXT_START) + p.content(p.until(TEXT_END)) + p.optional(p.literal(TEXT_END));
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + text_content + p.optional(p.literal(TURN_END)) + end;
+        }
+
+        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+        // <|START_ACTION|>[ {"tool_call_id": "0", "tool_name": "f", "parameters": {...}}, ... ]<|END_ACTION|>
+        auto tool_calls = p.standard_json_tools(ACTION_START, ACTION_END, inputs.tools, inputs.parallel_tool_calls,
+                                                /* force_tool_calls = */ true,
+                                                /* name_key         = */ "tool_name",
+                                                /* args_key         = */ "parameters",
+                                                /* array_wrapped    = */ true,
+                                                /* function_is_key  = */ false,
+                                                /* call_id_key      = */ "",
+                                                /* gen_call_id_key  = */ "tool_call_id",
+                                                /* parameters_order = */ { "tool_call_id", "tool_name", "parameters" });
+
+        // Content and tool calls are mutually exclusive in this format.
+        common_peg_parser body = require_tools ? tool_calls : p.choice({ tool_calls, text_content });
+
+        return generation_prompt + reasoning + body + p.optional(p.literal(TURN_END)) + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !has_response_format && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, ACTION_START }
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,287 @@
+#include "parsers.h"
+
+// The DeepSeek V4 reference implementation renders consecutive tool results into a single
+// user block, ordered by the tool call order of the preceding assistant message (matched
+// by tool call id) rather than by the order they appear in the conversation.
+static json deepseek_v4_sort_tool_results(const json & messages) {
+    json adjusted = messages;
+    std::map<std::string, size_t> call_order;
+
+    for (size_t i = 0; i < adjusted.size();) {
+        const auto & msg  = adjusted[i];
+        const auto   role = msg.value("role", "");
+
+        if (role == "assistant" && msg.contains("tool_calls") &&
+                msg.at("tool_calls").is_array() && !msg.at("tool_calls").empty()) {
+            call_order.clear();
+            const auto & tool_calls = msg.at("tool_calls");
+            for (size_t idx = 0; idx < tool_calls.size(); idx++) {
+                auto id = tool_calls[idx].value("id", "");
+                if (!id.empty()) {
+                    call_order[id] = idx;
+                }
+            }
+            i++;
+            continue;
+        }
+
+        if (role != "user" && role != "tool") {
+            i++;
+            continue;
+        }
+
+        // collect a maximal run of user/tool messages - they render into one user block
+        std::vector<size_t> tool_positions;
+        size_t run_end = i;
+        for (; run_end < adjusted.size(); run_end++) {
+            const auto r = adjusted[run_end].value("role", "");
+            if (r == "tool") {
+                tool_positions.push_back(run_end);
+            } else if (r != "user") {
+                break;
+            }
+        }
+
+        if (tool_positions.size() > 1 && !call_order.empty()) {
+            std::vector<json> results;
+            results.reserve(tool_positions.size());
+            for (auto pos : tool_positions) {
+                results.push_back(adjusted[pos]);
+            }
+            std::stable_sort(results.begin(), results.end(), [&](const json & a, const json & b) {
+                const auto order = [&](const json & m) {
+                    auto it = call_order.find(m.value("tool_call_id", ""));
+                    return it == call_order.end() ? (size_t) 0 : it->second;
+                };
+                return order(a) < order(b);
+            });
+            for (size_t k = 0; k < tool_positions.size(); k++) {
+                adjusted[tool_positions[k]] = std::move(results[k]);
+            }
+        }
+
+        i = run_end;
+    }
+
+    return adjusted;
+}
+
+common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_template &    tmpl,
+                                                                 const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    // V4 uses the same DSML markup as V3.2, but names the tool call block "tool_calls"
+    // instead of "function_calls", renders tool results in tool call order and its
+    // non-thinking generation prompt ends with a bare </think> instead of an empty
+    // <think></think> pair.
+    const bool is_v4 = tmpl.source().find("function_calls") == std::string::npos;
+
+    std::optional<json> adjusted_messages;
+    if (is_v4) {
+        adjusted_messages = deepseek_v4_sort_tool_results(inputs.messages);
+    }
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    std::optional<json> additional_context;
+    if (is_v4 && has_response_format) {
+        additional_context = json{ { "response_format", inputs.json_schema } };
+    }
+
+    const std::string DSML         = "｜DSML｜";
+    const std::string THINK_START  = "<think>";
+    const std::string THINK_END    = "</think>";
+    const std::string TC_BLOCK     = is_v4 ? "tool_calls" : "function_calls";
+    const std::string FC_START     = "<" + DSML + TC_BLOCK + ">";
+    const std::string FC_END       = "</" + DSML + TC_BLOCK + ">";
+    const std::string INVOKE_START = "<" + DSML + "invoke";
+    const std::string INVOKE_END   = "</" + DSML + "invoke>";
+    const std::string PARAM_START  = "<" + DSML + "parameter";
+    const std::string PARAM_END    = "</" + DSML + "parameter>";
+    const std::string GEN_PROMPT   = "<｜Assistant｜>";
+    const std::string TC_SEPARATOR = "\n\n";
+
+    data.prompt = common_chat_template_direct_apply_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = {THINK_END, FC_START};
+    data.preserved_tokens   = {
+        DSML,
+        THINK_START,
+        THINK_END,
+    };
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        if (is_v4 && msg.reasoning_content.empty()) {
+            data.generation_prompt = GEN_PROMPT + THINK_END;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += msg.render_content();
+            }
+        } else {
+            data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += THINK_END + msg.render_content();
+            }
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    bool require_tools   = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+    bool has_tool_calls = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal(GEN_PROMPT);
+        auto end               = p.end();
+
+        // build tool call section first since we might need it in reasoning
+        auto tool_choice = p.choice();
+        if (has_tool_calls) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
+                auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
+                const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
+
+                std::set<std::string> required;
+                if (params.contains("required")) {
+                    required = params.at("required").get<std::set<std::string>>();
+                }
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(params);
+
+                std::vector<common_peg_parser> required_parsers;
+                std::vector<common_peg_parser> optional_parsers;
+                for (const auto & [param_name, param_schema] : props.items()) {
+                    bool is_required = required.find(param_name) != required.end();
+                    bool is_string   = schema_info.resolves_to_string(param_schema);
+
+                    auto arg = p.tool_arg(
+                        p.tool_arg_open(p.literal(PARAM_START + " name=\"") + p.tool_arg_name(p.literal(param_name)) +
+                                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
+                        (is_string ?
+                             p.tool_arg_string_value(p.until(PARAM_END)) :
+                             p.tool_arg_json_value(p.schema(p.json(), "tool-" + name + "-arg-" + param_name + "-schema",
+                                                            param_schema, false))) +
+                        p.tool_arg_close(p.literal(PARAM_END)));
+
+                    auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
+                    if (is_required) {
+                        required_parsers.push_back(named_arg);
+                    } else {
+                        optional_parsers.push_back(named_arg);
+                    }
+                }
+
+                common_peg_parser args_seq = p.eps();
+                for (size_t i = 0; i < required_parsers.size(); i++) {
+                    if (i > 0) {
+                        args_seq = args_seq + p.space();
+                    }
+                    args_seq = args_seq + required_parsers[i];
+                }
+
+                if (!optional_parsers.empty()) {
+                    common_peg_parser any_opt = p.choice();
+                    for (const auto & opt : optional_parsers) {
+                        any_opt |= opt;
+                    }
+                    args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
+                }
+
+                common_peg_parser invoke_body = args_seq;
+                auto              func_parser = p.tool(p.tool_open(p.literal(INVOKE_START + " name=\"") +
+                                                                   p.tool_name(p.literal(name)) + p.literal("\">\n")) +
+                                                       invoke_body + p.space() + p.tool_close(p.literal(INVOKE_END)));
+
+                tool_choice |= p.rule("tool-" + name, func_parser);
+            });
+        }
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }
+
+        auto reasoning = p.eps();
+        auto reasoning_with_tc = p.eps();
+        auto obligatory_tool_calls = tool_calls;
+        bool allow_reasoning_with_tc = false;
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
+        if (extract_reasoning && inputs.enable_thinking) {
+            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+            reasoning_with_tc = THINK_START +
+                p.reasoning(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START, THINK_END })) +
+                p.space() + obligatory_tool_calls;
+            allow_reasoning_with_tc = true;
+        } else if (extract_reasoning) {
+            // Thinking disabled but reasoning extraction requested: the generation prompt
+            // contains an empty <think></think> pair (V3.2) or a bare </think> (V4) that
+            // must still be consumed.
+            reasoning = is_v4
+                ? p.optional(p.literal(THINK_END))
+                : p.optional(p.literal(THINK_START) + p.until(THINK_END) + p.literal(THINK_END));
+        }
+
+        if (has_response_format) {
+            auto response_format = p.rule("response-format",
+                p.literal("```json") + p.space() +
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
+                p.space() + p.literal("```"));
+            return generation_prompt + reasoning + response_format + end;
+        }
+
+        if (!has_tool_calls) {
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+
+        auto content_before_tools = p.negate(p.literal(THINK_START)) +
+            p.content(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START })) +
+            p.space();
+        return allow_reasoning_with_tc ? generation_prompt + (reasoning_with_tc | (reasoning + content_before_tools + tool_calls)) + end :
+            generation_prompt + reasoning + content_before_tools + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && !require_tools;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, FC_START },
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,101 @@
+#include "parsers.h"
+
+// Functionary v3.2 - uses recipient-based format: >>>recipient\n{content}
+common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template &    tmpl,
+                                                                   const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens  = {
+        ">>>all",
+    };
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+        data.generation_prompt = "<|start_header_id|>assistant<|end_header_id|>\n\n>>>all\n" + msg.render_content();
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        // Functionary v3.2 format:
+        // - Normal content: >>>all\n{content}
+        // - Tool calls: >>>function_name\n{json_args}
+        // Generation prompt ends with ">>>" so model outputs recipient immediately
+
+        // Build content parser for >>>all\n{content}
+        // When tools are present, content stops before the next ">>>" (tool call)
+        // When no tools, content goes until end
+        auto content_until_tool = p.literal("all\n") + p.content(p.until(">>>"));
+        auto content_until_end  = p.literal("all\n") + p.content(p.rest());
+        auto generation_prompt  = p.literal("<|start_header_id|>assistant<|end_header_id|>\n\n>>>");
+
+        // If no tools or tool_choice is NONE, just parse content
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            // When no tools, just match the prefix and capture everything after
+            return generation_prompt + content_until_end + p.end();
+        }
+
+        // Build tool call parsers for each available function
+        auto tool_choice = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            const auto & schema   = function.at("parameters");
+
+            // Tool format: >>>function_name\n{json_args}
+            auto tool_parser = p.tool(
+                p.tool_open(p.tool_name(p.literal(name)) + p.literal("\n")) +
+                p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema))
+            );
+
+            tool_choice |= p.rule("tool-" + name, tool_parser);
+        });
+
+        auto content_only = content_until_end;
+        auto tools_only = p.trigger_rule("tools", p.one_or_more(tool_choice));
+        auto content_and_tools = content_until_tool + tools_only;
+
+        auto ret = p.eps();
+        if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
+            if (inputs.parallel_tool_calls) {
+                ret = p.choice({ content_and_tools, tools_only }) + p.end();
+            } else {
+                ret = p.choice({ content_until_tool + tool_choice, tools_only }) + p.end();
+            }
+        } else if (inputs.parallel_tool_calls) {
+            ret = p.choice({ content_and_tools, content_only, tools_only }) + p.end();
+        } else {
+            auto content_and_tool = content_until_tool + tool_choice;
+            ret = p.choice({ content_and_tool, content_only, tool_choice }) + p.end();
+        }
+        return generation_prompt + ret;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        // Grammar trigger for when the model starts outputting a tool call
+        // (after the initial ">>>" in the generation prompt but recipient other than "all")
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, ">>>(?!all)" }
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,312 @@
+#include "parsers.h"
+
+namespace workaround {
+
+// Gemma4 uses a custom tool_responses field instead of role:tool messages.
+//
+// This will transform a sequence of messages:
+//   assistant(tool_call+) -> tool+ -> assistant(content)
+//
+// Into a single assistant message containing a tool_responses field:
+//   assistant(content + tool_call + tool_responses)
+//
+// This is necessary for the Gemma4 chat template to properly format the prompt.
+// See https://ai.google.dev/gemma/docs/core/prompt-formatting-gemma4
+struct gemma4_model_turn_builder {
+    json & messages;
+    size_t pos;
+    json tool_calls = json::array();
+    json tool_responses = json::array();
+    json content;
+    json reasoning_content;
+
+    gemma4_model_turn_builder(json & msgs, size_t pos) : messages(msgs), pos(pos) {}
+
+    void collect() {
+        // Collect the first assistant message
+        auto & msg = messages[pos];
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            // According to the prompt formatting guide, we need to preserve reasoning_content
+            // between function calls. The current chat templates do not support this, but we will do it anyway.
+            reasoning_content = msg.at("reasoning_content");
+        }
+        for (auto & tc : msg.at("tool_calls")) {
+            tool_calls.push_back(tc);
+        }
+        pos++;
+
+        // Collect tool call results
+        while (pos < messages.size() && messages[pos].value("role", "") == "tool") {
+            collect_result(messages[pos]);
+            pos++;
+        }
+
+        // Check if the next assistant message is the final message
+        if (pos < messages.size() && messages[pos].value("role", "") == "assistant") {
+            auto & next = messages[pos];
+            if (!has_tool_calls(next) && has_content(next)) {
+                content = next.at("content");
+                pos++;
+            }
+        }
+    }
+
+    void collect_result(const json & curr) {
+        json response;
+        if (curr.contains("content")) {
+            const auto & content = curr.at("content");
+            if (content.is_string()) {
+                // Try to parse the content as JSON; fall back to raw string
+                try {
+                    response = json::parse(content.get<std::string>());
+                } catch (...) {
+                    response = content;
+                }
+            } else {
+                response = content;
+            }
+        }
+
+        std::string name;
+
+        // Match name with corresponding tool call
+        size_t idx = tool_responses.size();
+        if (idx < tool_calls.size()) {
+            auto & tc = tool_calls[idx];
+            if (tc.contains("function")) {
+                name = tc.at("function").value("name", "");
+            }
+        }
+
+        // Fallback to the tool call id
+        if (name.empty()) {
+            name = curr.value("tool_call_id", "");
+        }
+
+        tool_responses.push_back({{"name", name}, {"response", response}});
+    }
+
+    json build() {
+        collect();
+
+        json msg = {
+            {"role", "assistant"},
+            {"tool_calls", tool_calls},
+        };
+        if (!tool_responses.empty()) {
+            msg["tool_responses"] = tool_responses;
+        }
+        if (!content.is_null()) {
+            msg["content"] = content;
+        }
+        if (!reasoning_content.is_null()) {
+            msg["reasoning_content"] = reasoning_content;
+        }
+        return msg;
+    }
+
+    static bool has_content(const json & msg) {
+        if (!msg.contains("content") || msg.at("content").is_null()) {
+            return false;
+        }
+        const auto & content = msg.at("content");
+        if (content.is_string() && !content.get<std::string>().empty()) {
+            return true;
+        }
+        if (content.is_array() && !content.empty()) {
+            return true;
+        }
+        return false;
+    }
+
+    static bool has_tool_calls(const json & msg) {
+        return msg.contains("tool_calls") && msg.at("tool_calls").is_array() && !msg.at("tool_calls").empty();
+    }
+};
+
+void convert_tool_responses_gemma4(json & messages) {
+    json result = json::array();
+    size_t i = 0;
+
+    while (i < messages.size()) {
+        auto & msg = messages[i];
+
+        if (msg.value("role", "") != "assistant" || !msg.contains("tool_calls") ||
+            !msg.at("tool_calls").is_array() || msg.at("tool_calls").empty()) {
+            result.push_back(msg);
+            i++;
+            continue;
+        }
+
+        gemma4_model_turn_builder builder(messages, i);
+        result.push_back(builder.build());
+        i = builder.pos;
+    }
+
+    messages = result;
+}
+
+}
+
+common_chat_params common_chat_params_init_gemma4(const common_chat_template &    tmpl,
+                                                         const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+
+    if (inputs.add_generation_prompt && string_ends_with(data.prompt, "<turn|>\n")) {
+        // This may happen if the model generates content + tool_call, the
+        // template does not add the model's next turn and confuses the model
+        // from emitting its proper reasoning token sequence.
+        data.generation_prompt = "<|turn>model\n";
+        data.prompt += data.generation_prompt;
+    }
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_USER,      "<|turn>user"  },
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|turn>model" },
+    };
+
+    data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "<|channel>thought";
+    data.thinking_end_tags  = {"<channel|>"};
+
+    data.preserved_tokens = {
+        "<|channel>",
+        "<channel|>",
+        "<|tool_call>",
+        "<tool_call|>",
+        "<|turn>",
+    };
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = string_ends_with(data.prompt, "<turn|>\n") ? "<|turn>model\n" : "";
+        data.generation_prompt += "<|channel>thought\n" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "<channel|>" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto start = p.rule("start", p.optional(p.literal("<|turn>model\n")));
+
+        if (extract_reasoning) {
+            p.rule("thought", p.literal("<|channel>thought") + p.space() + p.reasoning(p.until("<channel|>")) + p.literal("<channel|>"));
+        } else {
+            p.rule("thought", p.content(p.literal("<|channel>thought") + p.space() + p.until("<channel|>") + p.literal("<channel|>")));
+        }
+
+        auto consume_empty_channels = p.gbnf(p.zero_or_more(p.literal("<|channel>") + p.negate(p.literal("thought"))), "");
+        auto thought = (p.peek(p.literal("<|channel>")) + consume_empty_channels + p.ref("thought")) | p.negate(p.literal("<|channel>"));
+
+        if (has_response_format) {
+            auto response_format = p.literal("```json") <<
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) <<
+                p.literal("```");
+            return start + p.optional(thought) + response_format;
+        }
+
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            // Gemma4 tool calling syntax
+            // Rules should match traversal logic in gemma4_to_json()
+            p.rule("gemma4-string-content", p.until("<|\"|>"));
+            p.rule("gemma4-string", p.literal("<|\"|>") + p.ref("gemma4-string-content") + p.literal("<|\"|>"));
+            p.rule("gemma4-bool", p.json_bool());
+            p.rule("gemma4-null", p.json_null());
+            p.rule("gemma4-number", p.json_number());
+            p.rule("gemma4-dict-key", p.rule("gemma4-dict-key-name", p.chars("[^:}]", 1, -1)) + p.literal(":"));
+            p.rule("gemma4-dict-kv", p.ref("gemma4-dict-key") + p.space() + p.ref("gemma4-value"));
+            p.rule("gemma4-dict", [&]() {
+                auto ws = p.space();
+                auto member = p.ref("gemma4-dict-kv");
+                auto members = p.sequence({member, p.zero_or_more(p.sequence({p.literal(","), ws, member}))});
+                return p.sequence({
+                    p.literal("{"), ws,
+                    p.choice({p.literal("}"), p.sequence({members, ws, p.literal("}")})})
+                });
+            });
+            p.rule("gemma4-array", [&]() {
+                auto ws = p.space();
+                auto value = p.ref("gemma4-value");
+                auto elements = p.sequence({value, p.zero_or_more(p.sequence({p.literal(","), ws, value}))});
+                return p.sequence({
+                    p.literal("["), ws,
+                    p.choice({p.literal("]"), p.sequence({elements, ws, p.literal("]")})})
+                });
+            });
+            p.rule("gemma4-value", [&]() {
+                return p.choice({
+                    p.ref("gemma4-string"), p.ref("gemma4-dict"), p.ref("gemma4-array"),
+                    p.ref("gemma4-number"), p.ref("gemma4-bool"), p.ref("gemma4-null")
+                });
+            });
+
+            auto tool_choice = p.choice();
+
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
+                // TODO @aldehir : need to extend json-schema-to-grammar to produce more than JSON rules
+                // const auto & params   = function.at("parameters");
+
+                tool_choice |= p.rule("tool-" + name, p.tool(p.sequence({
+                    p.tool_open(p.tool_name(p.literal(name)) + p.peek(p.literal("{"))),
+                    p.tool_args(p.ref("gemma4-dict")),
+                })));
+            });
+
+            auto tool_call = p.trigger_rule("tool-call", p.repeat(
+                "<|tool_call>call:" + tool_choice + "<tool_call|>",
+                /* min = */ inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0,
+                /* max = */ inputs.parallel_tool_calls ? -1 : 1
+            ));
+
+            auto scan_to_toolcall = p.rule("scan-to-toolcall", p.until("<|tool_call>"));
+            auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<channel|>", "<|tool_call>"})));
+            auto message = p.rule("message", thought + content);
+            return start + p.zero_or_more(message) + scan_to_toolcall + tool_call;
+        }
+
+        // Gemma 4 may emit an extra <|channel>thought\n<channel|> at the end of the content. It may
+        // also emit a single trailing <channel|> token. Consume all complete reasoning blocks and
+        // then stop at the first unmatched <channel|> token.
+        auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<channel|>"})));
+        auto message = p.rule("message", thought + content);
+        return start + p.one_or_more(message);
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool_call>" },
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,81 @@
+#include "parsers.h"
+
+common_chat_params common_chat_params_init_gigachat_v3(
+        const common_chat_template & tmpl,
+        const autoparser::generation_params & inputs) {
+
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = false;
+    data.preserved_tokens  = {
+        "<|message_sep|>\n\n",
+        "<|role_sep|>\n",
+    };
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+        data.generation_prompt = "assistant<|role_sep|>\n" + msg.render_content();
+        data.prompt += data.generation_prompt;
+    }
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+    const auto *tool_call_start_prefix = "<|message_sep|>\n\nfunction call<|role_sep|>\n";
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto ret = p.eps();
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            // Build a choice of all available tools
+            auto tool_choice = p.choice();
+            for (const auto & tool : inputs.tools) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                const auto & schema = function.at("parameters");
+
+                auto tool_name = p.json_member("name", "\"" + p.tool_name(p.literal(name)) + "\"");
+                auto tool_args = p.json_member("arguments", p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
+
+                auto tool_open = p.tool_open(p.literal("{") << tool_name);
+
+                tool_choice |= p.rule("tool-" + name, tool_open << "," << tool_args << "}");
+            }
+
+            // Define the tool call structure
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+            auto max_calls = 1; // parallel toolcalls are not supported
+            auto tool_call = p.rule("tool-call", p.literal(tool_call_start_prefix) + tool_choice);
+            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(tool_call, /* min = */ min_calls, /* max = */ max_calls));
+
+            ret = p.content(p.until("<|message_sep|>\n\n")) << tool_calls;
+        } else {
+            // Content only parser
+            include_grammar = false;
+            ret = p.content(p.rest());
+        }
+
+        return p.literal("assistant<|role_sep|>\n") + ret;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto schema = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, tool_call_start_prefix}
+        };
+    }
+    return data;
+}
@@ -0,0 +1,167 @@
+#include "parsers.h"
+
+common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &    tmpl,
+                                                          const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    // Copy reasoning to the "thinking" field as expected by the gpt-oss template
+    auto adjusted_messages = json::array();
+    for (auto msg : inputs.messages) {
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            msg["thinking"] = msg.at("reasoning_content");
+            if (msg.contains("tool_calls") && msg.at("tool_calls").is_array() && !msg.at("tool_calls").empty()) {
+                msg.erase("content");
+            }
+        }
+        adjusted_messages.push_back(msg);
+    }
+
+    auto prompt = common_chat_template_direct_apply_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
+
+    // Check if we need to replace the return token with end token during
+    // inference and without generation prompt. For more details see:
+    // https://github.com/ggml-org/llama.cpp/issues/15417
+    if (inputs.is_inference && !inputs.add_generation_prompt) {
+        static constexpr std::string_view return_token = "<|return|>";
+        static constexpr std::string_view end_token    = "<|end|>";
+        if (size_t pos = prompt.rfind(return_token); pos != std::string::npos) {
+            prompt.replace(pos, return_token.length(), end_token);
+        }
+    }
+
+    data.prompt            = prompt;
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
+        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>developer" },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
+        { COMMON_CHAT_ROLE_TOOL,      "<|start|>functions" },
+    };
+
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+
+    data.thinking_start_tag = "<|channel|>analysis<|message|>";
+    data.thinking_end_tags  = {"<|end|>"};
+
+    // These special tokens are required to parse properly, so we include them
+    // even if parse_tool_calls is false.
+    data.preserved_tokens = {
+        "<|channel|>", "<|constrain|>", "<|message|>", "<|start|>", "<|end|>",
+    };
+
+    // Adjust prompt for continuation
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "<|start|>assistant<|channel|>analysis<|message|>" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "<|end|><|start|>assistant<|channel|>final<|message|>" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto start           = p.rule("start", p.literal("<|start|>assistant"));
+        auto end             = p.rule("end", p.literal("<|end|>"));
+        auto content         = p.rule("message-content", p.until("<|end|>"));
+        auto channel         = p.literal("<|channel|>") + (p.literal("commentary") | p.literal("analysis"));
+        auto constrain_type  = p.chars("[A-Za-z0-9_-]", 1, -1);
+
+        // Occasionally, gpt-oss-20b will prefix channels with this commentary
+        auto stray_commentary = p.optional(p.literal("<|channel|>commentary") + p.optional(p.literal(" to=assistant")));
+        auto start_analysis = stray_commentary + p.literal("<|channel|>analysis<|message|>");
+
+        if (extract_reasoning) {
+            p.rule("analysis", start_analysis + p.reasoning(content) + end);
+        } else {
+            p.rule("analysis", p.content(start_analysis + content + end));
+        }
+
+        auto analysis = p.ref("analysis");
+        auto preamble = p.rule("preamble", p.literal("<|channel|>commentary<|message|>") + p.content(content) + end);
+        auto final_msg = p.rule("final", stray_commentary + p.literal("<|channel|>final<|message|>") + p.content(content));
+
+        // Consume any unsolicited tool calls, e.g. builtin functions
+        auto unsolicited = p.rule("unsolicited", p.atomic(p.optional(channel) + p.literal(" to=") + content + end));
+
+        auto any = p.rule("any", preamble | analysis);
+
+        if (has_response_format) {
+            auto constraint = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);
+            auto response_format = p.rule("response-format",
+                p.literal("<|channel|>final") + constraint + p.literal("<|message|>") +
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
+
+            return p.zero_or_more(start + analysis) + start + response_format;
+        }
+
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto tool_choice = p.choice();
+
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
+                const auto & params   = function.at("parameters");
+
+                auto func_name  = p.literal(" to=functions.") + p.tool_name(p.literal(name));
+                auto constraint = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);
+                auto args       = p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", params));
+
+                // recipient in role header
+                //   <|start|>assistant to=functions.NAME<|channel|>(commentary|analysis)[constraint]<|message|>ARGS
+                auto tool_in_role = p.tool(p.tool_open(func_name + channel + constraint + p.literal("<|message|>")) + args);
+
+                // recipient in channel header
+                //   <|channel|>(commentary|analysis) to=functions.NAME[constraint]<|message|>ARGS
+                auto tool_in_channel = p.tool(p.tool_open(channel + func_name + constraint + p.literal("<|message|>")) + args);
+
+                tool_choice |= p.rule("tool-" + name, tool_in_role | tool_in_channel);
+            });
+
+            auto tool_call  = p.trigger_rule("tool-call", tool_choice);
+
+            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
+                return p.zero_or_more(start + any) + start + tool_call;
+            }
+
+            return p.zero_or_more(start + any) + start + (tool_call | final_msg);
+        }
+
+        return p.zero_or_more(start + any) + start + (final_msg | unsolicited);
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^\\s+to$" },
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^<\\|channel\\|>(?:commentary|analysis)\\s+to=functions$" },
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>assistant(\\s+to)" },
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>assistant(<\\|channel\\|>(?:commentary|analysis)\\s+to)" }
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,133 @@
+#include "parsers.h"
+
+// Kimi K2 Thinking - uses unique tool call ID format: functions.<name>:<index>
+// The ID contains both the function name and an incrementing counter
+common_chat_params common_chat_params_init_kimi_k2(const common_chat_template &    tmpl,
+                                                          const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+    data.preserved_tokens  = {
+        "<|tool_calls_section_begin|>",
+        "<|tool_calls_section_end|>",
+        "<|tool_call_begin|>",
+        "<|tool_call_argument_begin|>",
+        "<|tool_call_end|>",
+        "<think>",
+        "</think>",
+    };
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    const std::string SECTION_BEGIN = "<|tool_calls_section_begin|>";
+    const std::string SECTION_END   = "<|tool_calls_section_end|>";
+    const std::string CALL_BEGIN    = "<|tool_call_begin|>";
+    const std::string ARGS_BEGIN    = "<|tool_call_argument_begin|>";
+    const std::string CALL_END      = "<|tool_call_end|>";
+
+    const std::string THINK_START = "<think>";
+    const std::string THINK_END   = "</think>";
+    const std::string GEN_PROMPT  = "<|im_assistant|>assistant<|im_middle|>";
+
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = {THINK_END};
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        // Kimi K2 Thinking format:
+        // - Reasoning: <think>{reasoning}</think>
+        // - Content: text after reasoning
+        // - Tool calls section:
+        //   <|tool_calls_section_begin|>
+        //   <|tool_call_begin|>functions.<name>:<index><|tool_call_argument_begin|>{json_args}<|tool_call_end|>
+        //   ...
+        //   <|tool_calls_section_end|>
+        // The ID format is: functions.<function_name>:<counter> where counter is 0, 1, 2, ...
+
+        // Tool call markers
+        auto end = p.end();
+
+        // Note: this model is CRAZY. It can diverge from its supposed tool calling pattern in so many ways it's not funny.
+        // For example, it can call tools at the end of reasoning without closing reasoning...
+        auto reasoning = extract_reasoning ? p.optional(THINK_START + p.reasoning(
+            p.until_one_of({ THINK_END, "<|tool_calls_section_begin|>", "<|tool_call_begin|>" })) +
+            p.optional(p.literal(THINK_END))) : p.eps();
+        auto generation_prompt = p.literal(GEN_PROMPT);
+
+
+        // Content only parser (no tools)
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+
+        // Build tool call parsers for each available function
+        // The ID format is: functions.<name>:<index>
+        // We need to match: functions.<name>:<digits>
+        auto tool_choice = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            const auto & schema   = function.at("parameters");
+
+            // Match: functions.<name>:<digits>
+            // Capture the full call id (functions.<name>:<digits>) using tool_id tag
+            auto tool_id = p.tool_id(p.literal("functions.") + p.tool_name(p.literal(name)) + p.literal(":") + p.chars("[0-9]", 1, -1));
+            auto tool_parser = p.tool(
+                p.tool_open(tool_id + p.literal(ARGS_BEGIN)) +
+                p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)) +
+                p.tool_close(p.optional((p.literal(CALL_END))))
+            );
+
+            tool_choice |= p.rule("tool-" + name, tool_parser);
+        });
+
+        // Tool calls section: <|tool_calls_section_begin|> tool_calls <|tool_calls_section_end|>
+        auto min_calls  = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+        auto max_calls  = inputs.parallel_tool_calls ? -1 : 1;
+        // Use trigger_rule so grammar generator knows where to start generating rules
+        auto tool_calls = p.rule("tool-calls",
+            p.optional(p.literal(SECTION_BEGIN)) +
+            p.trigger_rule("tool-call", p.repeat(CALL_BEGIN + tool_choice, min_calls, max_calls) +
+                p.optional(p.literal(SECTION_END)))
+        );
+
+        auto content_before_tools = p.content(p.until_one_of({ SECTION_BEGIN, CALL_BEGIN }));
+
+        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool_call_begin|>" }
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,174 @@
+#include "parsers.h"
+
+// Kimi K3 - XTML tagged format, built by open_tag/close_tag macros:
+//   open_tag(t, attrs) = <|open|>t k="v"...<|sep|>   close_tag(t) = <|close|>t<|sep|>
+//   assistant := [think] [response] [tools] close_tag(message) <|end_of_msg|>
+// the generation prompt already opens the think (or response) section, so the
+// section opener is optional here - same as Kimi K2 Thinking
+common_chat_params common_chat_params_init_kimi_k3(const common_chat_template &          tmpl,
+                                                          const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+
+    const std::string SEP         = "<|sep|>";
+    const std::string MSG_START   = "<|open|>message role=\"assistant\"<|sep|>";
+    const std::string THINK_START = "<|open|>think<|sep|>";
+    const std::string THINK_END   = "<|close|>think<|sep|>";
+    const std::string RESP_START  = "<|open|>response<|sep|>";
+    const std::string RESP_END    = "<|close|>response<|sep|>";
+    const std::string TOOLS_START = "<|open|>tools<|sep|>";
+    const std::string TOOLS_END   = "<|close|>tools<|sep|>";
+    const std::string CALL_START  = "<|open|>call tool=\"";
+    const std::string CALL_END    = "<|close|>call<|sep|>";
+    const std::string ARG_START   = "<|open|>argument key=\"";
+    const std::string ARG_END     = "<|close|>argument<|sep|>";
+    const std::string MSG_END     = "<|close|>message<|sep|>";
+    const std::string EOM_TOKEN   = "<|end_of_msg|>";
+
+    // only the markers are special tokens. tag names ("think", "response", ...) are
+    // normal tokens and must not be preserved, or prose with those words is broken
+    data.preserved_tokens = {
+        "<|open|>",
+        "<|close|>",
+        "<|sep|>",
+        "<|end_of_msg|>",
+    };
+
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = { THINK_END };
+
+    // per-role message-start delimiters. user/assistant messages only have the role
+    // attribute, so the full opener is used. system and tool messages have more
+    // attributes, so those delimiters stop after the closing quote of the role
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|open|>message role=\"assistant\"<|sep|>" },
+        { COMMON_CHAT_ROLE_USER,      "<|open|>message role=\"user\"<|sep|>"      },
+        { COMMON_CHAT_ROLE_TOOL,      "<|open|>message role=\"tool\""             },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|open|>message role=\"system\""           },
+    };
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = MSG_START + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + RESP_START + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto end = p.end();
+
+        auto start = p.optional(p.literal(MSG_START));
+
+        // the think section is always consumed, even with reasoning extraction off:
+        // the generation prompt ends with open_tag('think'), so it is always present.
+        // reasoning stops at its own closer, or at the response opener if the model
+        // skips the closer
+        auto think_body = extract_reasoning ? p.reasoning(p.until_one_of({ THINK_END, RESP_START })) :
+                                              p.content(p.until_one_of({ THINK_END, RESP_START }));
+
+        auto reasoning = p.optional(p.optional(p.literal(THINK_START)) + think_body +
+                                    p.optional(p.literal(THINK_END)));
+
+        // content runs to the response closer, or to the next section if truncated
+        auto response = p.optional(p.literal(RESP_START)) +
+                        p.content(p.until_one_of({ RESP_END, TOOLS_START, MSG_END })) +
+                        p.optional(p.literal(RESP_END));
+
+        // the EOG token after the message closer reaches the parser as text,
+        // so it must be consumed or the parse stays incomplete
+        auto trailer = p.optional(p.literal(MSG_END)) + p.optional(p.literal(EOM_TOKEN));
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return start + reasoning + response + trailer + end;
+        }
+
+        auto tool_choices = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            const json   schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+            // arguments come one tag per key, with the JSON type in a type="..."
+            // attribute. the type is taken from the tool schema instead, as it tells
+            // us if the value is JSON or a literal string
+            auto args = p.eps();
+            if (schema.contains("properties") && !schema.at("properties").empty()) {
+                auto arg_choices = p.choice();
+                for (const auto & prop : schema.at("properties").items()) {
+                    const std::string & key = prop.key();
+
+                    std::string type = "string";
+                    if (prop.value().is_object() && prop.value().contains("type") &&
+                        prop.value().at("type").is_string()) {
+                        type = prop.value().at("type").get<std::string>();
+                    }
+
+                    auto value = type == "string" ? p.tool_arg_string_value(p.until(ARG_END)) :
+                                                    p.tool_arg_value(p.until(ARG_END));
+
+                    // skip the trailing type="..." attribute: anything up to <|sep|>
+                    arg_choices |= p.rule("kimi-k3-arg-" + name + "-" + key,
+                                          p.tool_arg(p.tool_arg_open(p.literal(ARG_START)) +
+                                                     p.tool_arg_name(p.literal(key)) + p.literal("\"") +
+                                                     p.until(SEP) + p.literal(SEP) + value +
+                                                     p.tool_arg_close(p.literal(ARG_END))));
+                }
+                args = p.zero_or_more(arg_choices);
+            }
+
+            // skip the trailing index="N" attribute the same way
+            auto call = p.tool(p.tool_open(p.literal(CALL_START) + p.tool_name(p.literal(name)) + p.literal("\"") +
+                                           p.until(SEP) + p.literal(SEP)) +
+                               p.tool_args(args) + p.tool_close(p.literal(CALL_END)));
+
+            tool_choices |= p.rule("kimi-k3-tool-" + name, call);
+        });
+
+        // all calls go inside one tools section, then the message is closed. the
+        // message closer is part of the trigger rule, or else the lazy grammar
+        // rejects it once tool calls have started
+        auto tools_section =
+            p.trigger_rule("kimi-k3-tool-call", p.literal(TOOLS_START) + p.one_or_more(tool_choices) +
+                                                    p.literal(TOOLS_END) + p.optional(p.literal(MSG_END)) +
+                                                    p.optional(p.literal(EOM_TOKEN)));
+
+        auto tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? tools_section :
+                                                                              p.optional(tools_section);
+
+        return start + reasoning + response + tools + trailer + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                if (function.contains("parameters")) {
+                    auto schema = function.at("parameters");
+                    builder.resolve_refs(schema);
+                }
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOLS_START },
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,119 @@
+#include "parsers.h"
+
+// LFM2 format detection: template uses <|tool_list_start|>[...]<|tool_list_end|> around the tool list
+// and <|tool_call_start|>[...]<|tool_call_end|> around each tool call
+bool is_lfm2_template(const std::string & src) {
+    return src.find("<|tool_list_start|>") != std::string::npos &&
+           src.find("<|tool_list_end|>")   != std::string::npos;
+}
+
+// LFM2/LFM2.5 parser. Tool calls are almost Python-style and parallel-capable
+// (except dotted names and JSON literals true/false/null).
+// Always wrapped in <|tool_call_start|>[name(args)]<|tool_call_end|> with optional <think> reasoning.
+// tool_list_tokens preserves LFM2 system tool-list markers.
+common_chat_params common_chat_params_init_lfm2(const common_chat_template &          tmpl,
+                                                       const autoparser::generation_params & inputs,
+                                                       bool tool_list_tokens) {
+    common_chat_params data;
+
+    const std::string TOOL_CALL_START = "<|tool_call_start|>";
+    const std::string TOOL_CALL_END   = "<|tool_call_end|>";
+    const std::string TOOL_LIST_START = "<|tool_list_start|>";
+    const std::string TOOL_LIST_END   = "<|tool_list_end|>";
+    const std::string THINK_START     = "<think>";
+    const std::string THINK_END       = "</think>";
+    const std::string GEN_PROMPT      = "<|im_start|>assistant\n";
+
+    // Copy reasoning to the "thinking" field the template expects
+    auto adjusted_messages = json::array();
+    for (auto msg : inputs.messages) {
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            msg["thinking"] = msg.at("reasoning_content");
+        }
+        adjusted_messages.push_back(msg);
+    }
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+    data.preserved_tokens  = { TOOL_CALL_START, TOOL_CALL_END, THINK_START, THINK_END };
+    if (tool_list_tokens) {
+        data.preserved_tokens.push_back(TOOL_LIST_START);
+        data.preserved_tokens.push_back(TOOL_LIST_END);
+    }
+
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = {THINK_END};
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    // Gate by reasoning format and whether the template supports <think>
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE &&
+                             tmpl.source().find(THINK_START) != std::string::npos;
+    auto include_grammar   = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal(GEN_PROMPT);
+        auto end = p.end();
+
+        auto reasoning = p.eps();
+        if (extract_reasoning) {
+            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+        }
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            if (has_response_format) {
+                auto response_format = p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema));
+                return generation_prompt + reasoning + response_format + end;
+            }
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+        auto tool_calls = p.rule("tool-calls",
+            p.trigger_rule("tool-call",
+                p.literal(TOOL_CALL_START) +
+                p.python_style_tool_calls(inputs.tools, inputs.parallel_tool_calls, /* allow_json_literals = */ true) +
+                p.literal(TOOL_CALL_END)
+            )
+        );
+
+        auto content = p.content(p.until(TOOL_CALL_START));
+
+        return generation_prompt + reasoning + content + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOL_CALL_START }
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,144 @@
+#include "parsers.h"
+
+// MiniCPM5 format:
+// - Reasoning: <think>{reasoning}</think> (optional)
+// - Tool calls: <function name="foo"><param name="bar">value</param></function>
+common_chat_params common_chat_params_init_minicpm5(const common_chat_template &          tmpl,
+                                                           const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+    data.preserved_tokens  = {
+        "<function",
+        "<param",
+        "</function>",
+        "</param>",
+        "<think>",
+        "</think>",
+    };
+
+    data.thinking_start_tag = "<think>";
+    data.thinking_end_tags  = {"</think>"};
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|im_start|>assistant"             },
+        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>user\n<tool_response>" },
+        { COMMON_CHAT_ROLE_USER,      "<|im_start|>user"                  },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|im_start|>system"                },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "<|im_start|>assistant\n<think>\n" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "\n</think>\n\n" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal("<|im_start|>assistant\n");
+
+        auto reasoning = p.eps();
+        if (extract_reasoning) {
+            reasoning = ("<think>" << p.reasoning(p.until("</think>")) << "</think>") + p.space();
+        }
+
+        // Response format parser
+        if (has_response_format) {
+            return generation_prompt + reasoning + p.content(p.schema(p.json(), "response-format", inputs.json_schema));
+        }
+
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            // CDATA lets a value carry characters that would otherwise close the tag (e.g.
+            // </param>); capture the inner text only, excluding the CDATA markers.
+            auto string_value = p.choice({
+                p.literal("<![CDATA[") + p.ac(p.tool_arg_string_value(p.until("]]>")) + p.literal("]]>"), "]]>") + p.tool_arg_close(p.literal("</param>")),
+                p.negate(p.literal("<![CDATA[")) + p.ac(p.tool_arg_string_value(p.until("</param>")) + p.tool_arg_close(p.literal("</param>")), "</param>")
+            });
+
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto &      function = tool.at("function");
+                const std::string name     = function.at("name");
+                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                auto args = p.eps();
+                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
+                    auto schema_info = common_schema_info();
+                    schema_info.resolve_refs(params);
+
+                    auto arg_choice = p.choice();
+                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
+                        auto value_parser = p.eps();
+                        if (schema_info.resolves_to_string(prop_schema)) {
+                            value_parser = string_value;
+                        } else {
+                            value_parser = p.tool_arg_json_value(
+                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false)
+                                ) + p.tool_arg_close(p.literal("</param>"));
+                        }
+
+                        auto arg_rule = p.tool_arg(
+                            p.tool_arg_open(p.literal("<param name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
+                            value_parser
+                        );
+
+                        arg_choice |= arg_rule;
+                    }
+                    args = p.zero_or_more(arg_choice + p.space());
+                }
+
+                auto tool_parser = p.tool(
+                    p.tool_open(p.literal("<function name=\"") + p.tool_name(p.literal(name)) + p.literal("\">"))
+                    << p.tool_args(args)
+                    << p.tool_close(p.literal("</function>")));
+
+                tool_choice |= p.rule("tool-" + name, tool_parser);
+            });
+
+            auto max_calls  = inputs.parallel_tool_calls ? -1 : 1;
+            auto tool_calls = p.trigger_rule("tool-call", p.repeat(tool_choice + p.space(), 1, max_calls));
+
+            auto content = p.content(p.until("<function"));
+
+            return generation_prompt + reasoning + content + tool_calls + p.end();
+        }
+
+        return generation_prompt + reasoning + p.content(p.rest()) + p.end();
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function" },
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,259 @@
+#include "parsers.h"
+
+common_chat_params common_chat_params_init_minimax_m3(const common_chat_template &          tmpl,
+                                                             const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_MINIMAX_M3;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "<mm:think>";
+    data.thinking_end_tags  = {"</mm:think>"};
+
+    // M3 prefixes every tool tag with the namespace token "]<]minimax[>[";
+    // params use the parameter name as the tag (<file_path>...</file_path>).
+    const std::string NS          = "]<]minimax[>[";
+    const std::string THINK_START = "<mm:think>";
+    const std::string THINK_END   = "</mm:think>";
+    const std::string FC_START    = NS + "<tool_call>";
+    const std::string FC_END      = NS + "</tool_call>";
+    const std::string INVOKE_END  = NS + "</invoke>";
+
+    data.preserved_tokens = {
+        NS,
+        "<tool_call>",
+        "</tool_call>",
+        THINK_START,
+        THINK_END,
+    };
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "]~b]ai"        },
+        { COMMON_CHAT_ROLE_USER,      "]~b]user"      },
+        { COMMON_CHAT_ROLE_TOOL,      "]~b]tool"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]developer" },
+        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]system"    },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    const std::string GEN_PROMPT = data.generation_prompt;
+
+    using mm3 = common_chat_peg_minimax_m3_mapper;
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.prefix(GEN_PROMPT, THINK_START);
+        auto end = p.end();
+
+        auto reasoning = p.eps();
+        if (extract_reasoning) {
+            auto block = inputs.enable_thinking
+                             ? p.literal(THINK_START) + p.space() +
+                                   p.ac(p.reasoning(p.until(THINK_END)) + p.literal(THINK_END), THINK_END)
+                             : p.literal(THINK_START) + p.ac(p.until(THINK_END) + p.literal(THINK_END), THINK_END);
+
+            // A turn without reasoning is prefixed with a bare </mm:think>, written either by the
+            // generation prompt (thinking_mode = "disabled") or by the model itself.
+            reasoning = p.optional(p.choice({ block, p.literal(THINK_END) }));
+        }
+
+        if (has_response_format) {
+            auto response_format = p.rule("response-format",
+                p.literal("```json") + p.space() +
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
+                p.space() + p.literal("```"));
+            return generation_prompt + reasoning + response_format + end;
+        }
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+
+        auto alternatives_of = [](const json & schema) -> std::optional<json> {
+            for (const auto * keyword : { "oneOf", "anyOf" }) {
+                if (schema.contains(keyword) && schema.at(keyword).is_array() && !schema.at(keyword).empty()) {
+                    return schema.at(keyword);
+                }
+            }
+            return std::nullopt;
+        };
+
+        auto tool_choice = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+            auto schema_info = common_schema_info();
+            schema_info.resolve_refs(params);
+
+            // The template expands argument values recursively in XML (see the to_xml() macro)
+            std::function<common_peg_parser(const json &, const std::string &, const std::string &)> value_of;
+            std::function<common_peg_parser(const json &, const std::string &)>                      members_of;
+
+            auto element_of = [&](const std::string & tag, const json & schema, const std::string & rule_name) {
+                const std::string close = NS + "</" + tag + ">";
+                return p.rule(rule_name,
+                    p.tool_arg(
+                        p.tool_arg_open(
+                            p.literal(NS + "<") +
+                            p.tool_arg_name(p.literal(tag)) +
+                            p.literal(">")) +
+                        value_of(schema, rule_name, close)));
+            };
+
+            value_of = [&](const json & schema,
+                           const std::string & rule_name,
+                           const std::string & close) -> common_peg_parser {
+                auto close_tag = p.tool_arg_close(p.literal(close));
+
+                // A string accepts anything, so a union with a string alternative is a string
+                if (schema_info.resolves_to_string(schema)) {
+                    return p.ac(p.tool_arg_string_value(p.until(close)) + close_tag, close);
+                }
+
+                if (auto alternatives = alternatives_of(schema)) {
+                    std::vector<common_peg_parser> choices;
+
+                    size_t index = 0;
+                    for (const auto & alternative : *alternatives) {
+                        const std::string alt_name = rule_name + "-" + std::to_string(index++);
+
+                        // There is a risk that this breaks streaming deltas, but that's a risk we
+                        // assume to provide tool arg streaming.
+                        choices.push_back(value_of(alternative, alt_name, close));
+                    }
+
+                    return p.choice(choices);
+                }
+
+                const std::string type = schema.contains("type") && schema.at("type").is_string()
+                                             ? schema.at("type").get<std::string>()
+                                             : "";
+
+                if (type == "object" && schema.contains("properties")) {
+                    return p.tag(mm3::TOOL_ARG_OBJECT, members_of(schema, rule_name)) + p.space() + close_tag;
+                }
+
+                if (type == "array" && schema.contains("items")) {
+                    const std::string item_close = NS + "</item>";
+                    auto item = p.rule(rule_name + "-item",
+                        p.tag(mm3::TOOL_ARG_ITEM,
+                              p.literal(NS + "<item>") +
+                                  value_of(schema.at("items"), rule_name + "-item", item_close)));
+                    return p.tag(mm3::TOOL_ARG_ARRAY, p.repeat(p.space() + item, 0, -1)) + p.space() + close_tag;
+                }
+
+                return p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", schema, false)) + close_tag;
+            };
+
+            // Required properties in schema order, then any number of optional ones in any order.
+            members_of = [&](const json & schema, const std::string & rule_prefix) -> common_peg_parser {
+                const auto & props = schema.at("properties");
+
+                std::set<std::string> required;
+                if (schema.contains("required")) {
+                    required = schema.at("required").get<std::set<std::string>>();
+                }
+
+                std::vector<common_peg_parser> required_elements;
+                std::vector<common_peg_parser> optional_elements;
+                for (const auto & [key, key_schema] : props.items()) {
+                    auto element = element_of(key, key_schema, rule_prefix + "-" + key);
+                    if (required.find(key) != required.end()) {
+                        required_elements.push_back(element);
+                    } else {
+                        optional_elements.push_back(element);
+                    }
+                }
+
+                common_peg_parser members = p.eps();
+                for (size_t i = 0; i < required_elements.size(); i++) {
+                    if (i > 0) {
+                        members = members + p.space();
+                    }
+                    members = members + required_elements[i];
+                }
+
+                if (!optional_elements.empty()) {
+                    common_peg_parser any_optional = p.choice();
+                    for (const auto & element : optional_elements) {
+                        any_optional |= element;
+                    }
+                    members = members + p.repeat(p.space() + any_optional, 0, -1);
+                }
+
+                return members;
+            };
+
+            common_peg_parser invoke_body =
+                params.contains("properties") ? members_of(params, "tool-" + name + "-arg") : p.eps();
+
+            auto func_parser = p.tool(
+                p.tool_open(p.literal(NS + "<invoke name=\"") +
+                            p.tool_name(p.literal(name)) + p.literal("\">")) +
+                p.space() + invoke_body + p.space() +
+                p.tool_close(p.literal(INVOKE_END)));
+
+            tool_choice |= p.rule("tool-" + name, func_parser);
+        });
+
+        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
+        auto content_before_tools = p.content(p.until(FC_START));
+        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, FC_START },
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,135 @@
+#include "parsers.h"
+
+common_chat_params common_chat_params_init_ministral_3(const common_chat_template &    tmpl,
+                                                              const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    // Build up messages to follow the format: https://huggingface.co/mistralai/Ministral-3-14B-Reasoning-2512/blob/main/chat_template.jinja
+    auto adjusted_messages = json::array();
+    for (const auto & msg : inputs.messages) {
+        auto role = msg.value("role", "");
+        if (role != "system" && role != "assistant") {
+            // Only adjust system and assistant messages. Interestingly, the system message may contain thinking.
+            adjusted_messages.push_back(msg);
+            continue;
+        }
+
+        auto content = json::array();
+
+        // If message contains `reasoning_content`, add it as a block of type `thinking`
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            content.push_back({
+                { "type",     "thinking"                                     },
+                { "thinking", msg.at("reasoning_content").get<std::string>() },
+            });
+        }
+
+        // If message contains `content`, add it as a block of type `text`
+        if (msg.contains("content")) {
+            if (msg.at("content").is_string()) {
+                content.push_back({
+                    { "type", "text"                               },
+                    { "text", msg.at("content").get<std::string>() },
+                });
+            } else if (msg.at("content").is_array()) {
+                auto blocks = msg.at("content");
+                content.insert(blocks);
+            }
+        }
+
+        auto adjusted       = msg;
+        adjusted["content"] = content;
+        adjusted.erase("reasoning_content");
+        adjusted_messages.push_back(adjusted);
+    }
+
+    auto has_tools            = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format  = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+    auto extract_reasoning    = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar      = true;
+
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "[THINK]";
+    data.thinking_end_tags  = {"[/THINK]"};
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens  = {
+        "[THINK]",
+        "[/THINK]",
+        "[TOOL_CALLS]",
+        "[ARGS]",
+    };
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "[THINK]" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "[/THINK]" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.eps();
+        auto reasoning =
+            extract_reasoning ? p.optional("[THINK]" + p.reasoning(p.until("[/THINK]")) + "[/THINK]") : p.eps();
+
+        // Response format parser
+        if (has_response_format) {
+            // Ministral wants to emit json surrounded by code fences
+            return generation_prompt + (reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```");
+        }
+
+        // Tool call parser
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
+                const auto & schema   = function.at("parameters");
+
+                tool_choice |=
+                    p.rule("tool-" + name, p.tool_open(p.tool_name(p.literal(name)) + "[ARGS]") +
+                                               p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
+            });
+
+            auto min_calls  = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+            auto max_calls  = inputs.parallel_tool_calls ? -1 : 1;
+            auto tool_calls = p.trigger_rule("tool-call", p.repeat("[TOOL_CALLS]" + tool_choice, min_calls, max_calls));
+
+            return generation_prompt + (reasoning << p.content(p.until("[TOOL_CALLS]")) << tool_calls);
+        }
+
+        // Content only parser
+        include_grammar = false;
+        return generation_prompt + (reasoning << p.content(p.rest()));
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]" }
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,148 @@
+#include "parsers.h"
+
+// An assistant turn is rendered as one or more messages, each
+// "<|start|>assistant to=<recipient><|message|>{content}{END}" where END is
+// <|eom|> (more messages follow) or <|eot|> (end of turn):
+//   - chain-of-thought: to=self, terminated by <|eom|>
+//   - final answer:     to=user, terminated by <|eot|>
+// The generation prompt is just "<|start|>assistant"; the model emits its own
+// " to=...<|message|>".
+common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template &          tmpl,
+                                                               const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = "<|start|>assistant";
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+
+    data.preserved_tokens = {
+        "<|start|>", "<|message|>", "<|eom|>", "<|eot|>",
+        // ATEM tool-call markup emitted on " to=<tool>" turns.
+        "<atem:function_calls>", "<atem:invoke", "<atem:parameter", "</atem:parameter>",
+        "</atem:invoke>", "</atem:function_calls>",
+    };
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
+        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
+        { COMMON_CHAT_ROLE_TOOL,      "<|start|>tool"      },
+    };
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = "<|start|>assistant to=self<|message|>" + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += "<|eom|><|start|>assistant to=user<|message|>" + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+
+    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+    // Constrained grammar whenever tools are offered.
+    auto include_grammar = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto start = p.rule("start", p.literal("<|start|>assistant"));
+
+        if (!extract_reasoning && !include_grammar) {
+            return start + p.content(p.rest());
+        }
+
+        if (extract_reasoning) {
+            p.rule("analysis", p.literal(" to=self<|message|>") + p.reasoning(p.until("<|eom|>")) + p.literal("<|eom|>"));
+        } else {
+            p.rule("analysis", p.literal(" to=self<|message|>") + p.content(p.until("<|eom|>")) + p.literal("<|eom|>"));
+        }
+        auto analysis = p.ref("analysis");
+
+        auto recipient  = p.optional(p.literal(" to=user"));
+        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") +
+                                              p.content(p.until_one_of({ "<|eot|>", "<|eom|>" })));
+
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto string_value = p.ac(
+                p.tool_arg_string_value(p.until("</atem:parameter>")) + p.tool_arg_close(p.literal("</atem:parameter>")),
+                "</atem:parameter>");
+
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto &      function = tool.at("function");
+                const std::string name     = function.at("name");
+                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                auto args = p.eps();
+                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
+                    auto schema_info = common_schema_info();
+                    schema_info.resolve_refs(params);
+
+                    auto arg_choice = p.choice();
+                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
+                        auto value_parser = p.eps();
+                        if (schema_info.resolves_to_string(prop_schema)) {
+                            value_parser = string_value;
+                        } else {
+                            value_parser = p.tool_arg_json_value(
+                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false))
+                                + p.tool_arg_close(p.literal("</atem:parameter>"));
+                        }
+
+                        auto arg_rule = p.tool_arg(
+                            p.tool_arg_open(p.literal("<atem:parameter name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
+                            value_parser);
+
+                        arg_choice |= arg_rule;
+                    }
+                    args = p.zero_or_more(arg_choice + p.space());
+                }
+
+                auto tool_parser = p.tool(
+                    p.tool_open(p.literal(" to=") + p.until("<|message|>") +
+                                p.literal("<|message|><atem:function_calls>") + p.space() +
+                                p.literal("<atem:invoke name=\"") + p.tool_name(p.literal(name)) + p.literal("\">") + p.space())
+                    << p.tool_args(args)
+                    << p.tool_close(p.literal("</atem:invoke>") + p.space() + p.literal("</atem:function_calls>")));
+
+                tool_choice |= p.rule("tool-" + name, tool_parser);
+            });
+
+            auto tool_calls = inputs.parallel_tool_calls
+                ? p.trigger_rule("tool-call", tool_choice + p.zero_or_more(p.literal("<|eom|>") + start + tool_choice))
+                : p.trigger_rule("tool-call", tool_choice);
+
+
+            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
+                return p.zero_or_more(start + analysis) + start + tool_calls;
+            }
+            auto trailing_calls = p.optional(p.literal("<|eom|>") + start + tool_calls);
+            return p.zero_or_more(start + analysis) + start + (tool_calls | (final_msg + trailing_calls));
+        }
+
+        return p.zero_or_more(start + analysis) + start + final_msg;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+              "<\\|start\\|>assistant( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
+        };
+    }
+
+    return data;
+}
@@ -0,0 +1,34 @@
+#include "parsers.h"
+
+#include "log.h"
+
+#include <set>
+
+void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
+    for (const auto & tool : tools) {
+        if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
+            LOG_INF("Skipping tool without function: %s", tool.dump(2).c_str());
+            continue;
+        }
+        fn(tool);
+    }
+}
+
+void foreach_parameter(const json & function, const std::function<void(const std::string &, const json &, bool)> & fn) {
+    if (!function.contains("parameters") || !function.at("parameters").is_object()) {
+        return;
+    }
+    const auto & params = function.at("parameters");
+    if (!params.contains("properties") || !params.at("properties").is_object()) {
+        return;
+    }
+    const auto & props = params.at("properties");
+    std::set<std::string> required;
+    if (params.contains("required") && params.at("required").is_array()) {
+        required = params.at("required").get<std::set<std::string>>();
+    }
+    for (const auto & [name, prop] : props.items()) {
+        bool is_required = (required.find(name) != required.end());
+        fn(name, prop, is_required);
+    }
+}
@@ -0,0 +1,77 @@
+#pragma once
+
+#include "chat.h"
+#include "chat-auto-parser.h"
+#include "chat-auto-parser-helpers.h"
+#include "chat-peg-parser.h"
+#include "common.h"
+#include "ggml.h"
+#include "json-schema-to-grammar.h"
+#include "json.h"
+
+#include <functional>
+#include <optional>
+#include <set>
+#include <string>
+#include <vector>
+
+using json = common_json;
+
+// iterate over the function tools of an OpenAI-style tools array
+void foreach_function(const json & tools, const std::function<void(const json &)> & fn);
+
+// iterate over the parameters of a function tool, flagging the ones listed as required
+void foreach_parameter(const json & function, const std::function<void(const std::string &, const json &, bool)> & fn);
+
+// render a template; the override arguments let a parser feed in messages, tools or context it has rewritten
+std::string common_chat_template_direct_apply_impl(
+    const common_chat_template & tmpl,
+    const autoparser::generation_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt);
+
+// the suffix a template appends when add_generation_prompt is set
+std::string common_chat_template_generation_prompt_impl(
+    const common_chat_template & tmpl,
+    const autoparser::generation_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt);
+
+bool is_lfm2_template(const std::string & src);
+
+namespace workaround {
+
+void convert_tool_responses_gemma4(json & messages);
+
+}
+
+common_chat_params common_chat_params_init_cohere2moe(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_gemma4(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_gigachat_v3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_kimi_k2(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_kimi_k3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+// tool_list_tokens preserves the LFM2 system tool-list markers; LFM2.5 renders without them
+common_chat_params common_chat_params_init_lfm2(const common_chat_template & tmpl, const autoparser::generation_params & inputs, bool tool_list_tokens);
+
+common_chat_params common_chat_params_init_minicpm5(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_minimax_m3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_ministral_3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
+
+common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
@@ -0,0 +1,181 @@
+#include "parsers.h"
+
+common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template &          tmpl,
+                                                              const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    const std::string GEN_PREFIX = "<|im_start|>assistant\n";
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+
+    auto supports_reasoning = tmpl.source().find("<think>") != std::string::npos;
+
+    data.supports_thinking = supports_reasoning;
+    data.preserved_tokens  = {
+        "<tool_call>",
+        "</tool_call>",
+    };
+
+    auto is_qwen3_coder  = !supports_reasoning;
+
+    if (supports_reasoning) {
+        data.thinking_start_tag = "<think>";
+        // Support both </think> and <tool_call> as reasoning end sequences.
+        // <function= is omitted, as it is a workaround for Qwen3-Coder which is not a thinking model
+        data.thinking_end_tags = { "</think>", "<tool_call>" };
+        data.preserved_tokens.insert(data.preserved_tokens.end(), { "<think>", "</think>" });
+    }
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|im_start|>assistant"             },
+        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>user\n<tool_response>" }, // Qwen3-Coder, Qwen3.5, Nemotron Nano 3
+        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>tool_response"         }, // StepFun-3.5-Flash
+        { COMMON_CHAT_ROLE_USER,      "<|im_start|>user"                  },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|im_start|>system"                },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PREFIX;
+        if (supports_reasoning) {
+            data.generation_prompt += "<think>\n" + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += "\n</think>\n\n";
+            }
+        }
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    std::vector<std::string> tool_call_starts = { "<tool_call>" };
+
+    if (is_qwen3_coder) {
+        // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
+        // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
+        // constraining on <function which may occur in valid content generation, e.g. #include <functional>
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const std::string name = tool.at("function").at("name");
+            tool_call_starts.push_back("<function=" + name + ">");
+        });
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal(GEN_PREFIX);
+
+        auto reasoning = p.eps();
+        if (supports_reasoning && extract_reasoning) {
+            reasoning = p.optional("<think>" + p.space() +
+                                   p.reasoning(p.until_one_of({ "</think>", "<tool_call>" })) +
+                                   (p.literal("</think>") | p.peek(p.literal("<tool_call>"))));
+        }
+
+        // Response format parser
+        if (has_response_format) {
+            return generation_prompt + (reasoning << p.content(p.schema(p.json(), "response-format", inputs.json_schema)));
+        }
+
+        // Tool call parser
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto arg_close  = p.tool_arg_close(p.literal("\n</parameter>\n"));
+            auto arg_string = p.rule("xml-arg-string",
+                p.ac(p.tool_arg_string_value(p.until("\n</parameter>\n")) + arg_close, "\n</parameter>\n"));
+
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function   = tool.at("function");
+                std::string  name       = function.at("name");
+                auto         parameters = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(parameters);
+
+                std::vector<common_peg_parser> required_args;
+                std::vector<common_peg_parser> optional_args;
+
+                foreach_parameter(function, [&](const std::string & param_name, const json & param_schema, bool is_required) {
+                    auto rule_name = "tool-" + name + "-arg-" + param_name;
+
+                    auto arg_open = p.tool_arg_open("<parameter=" + p.tool_arg_name(p.literal(param_name)) + ">\n");
+
+                    auto arg_value = schema_info.resolves_to_string(param_schema) ?
+                        arg_string :
+                        p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", param_schema)) + arg_close;
+
+                    auto arg_rule = p.rule(rule_name, p.tool_arg(arg_open + arg_value));
+
+                    (is_required ? required_args : optional_args).push_back(arg_rule);
+                });
+
+                // Accept required arguments in any order, as Qwen does not always adhere to the
+                // order provided.
+                auto args = p.permute("tool-" + name + "-args", required_args);
+                if (!optional_args.empty()) {
+                    args = args + p.zero_or_more(p.choice(optional_args));
+                }
+
+                auto func = p.tool(p.tool_open("<function=" + p.tool_name(p.literal(name)) + ">\n") +
+                                   p.tool_args(args) +
+                                   p.tool_close(p.literal("</function>\n")));
+
+                tool_choice |= p.rule("tool-" + name, func);
+            });
+
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+
+            auto tool_call_body = tool_choice + "</tool_call>" + p.space();
+            auto tool_call      = p.rule("tool-call", "<tool_call>\n" + tool_call_body);
+
+            // Qwen3-Coder models may occasionally omit the <tool_call> token.
+            auto tool_call_first = is_qwen3_coder ?
+                p.rule("tool-call-first", p.optional(p.literal("<tool_call>\n")) + tool_call_body) :
+                tool_call;
+
+            auto calls      = inputs.parallel_tool_calls ? tool_call_first + p.zero_or_more(tool_call) : tool_call_first;
+            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));
+
+            return generation_prompt +
+                   (reasoning << p.content(p.until_one_of(tool_call_starts)) << tool_calls);
+        }
+
+        // Content only parser
+        return generation_prompt + (reasoning << p.content(p.rest()));
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        if (data.grammar_lazy) {
+            for (const auto & start : tool_call_starts) {
+                data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, start });
+            }
+        }
+    }
+
+    return data;
+}
@@ -0,0 +1,20 @@
+# Specialized chat template parsers, listed explicitly so that adding or removing one re-runs CMake instead of leaving an incremental build stale.
+
+set(LLAMA_CHAT_PARSERS_SOURCES
+    ${CMAKE_CURRENT_LIST_DIR}/parsers.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/parsers.h
+    ${CMAKE_CURRENT_LIST_DIR}/cohere2moe.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/deepseek.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/functionary-v3-2.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/gemma4.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/gigachat-v3.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/gpt-oss.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/kimi-k2.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/kimi-k3.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/lfm2.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/minicpm5.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/minimax-m3.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/ministral3.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/muse-glimmer.cpp
+    ${CMAKE_CURRENT_LIST_DIR}/qwen3-coder.cpp
+)
@@ -2467,11 +2467,22 @@ common_params common_base_params_to_speculative(const common_params & params) {
    result.pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED;

    if (has_draft) {
-        result.devices               = params_spec.devices;
+        // default to global devices value
+        if (!params_spec.devices.empty()) {
+            result.devices           = params_spec.devices;
+        }
        result.model                 = params_spec.mparams;
        result.n_gpu_layers          = params_spec.n_gpu_layers;
        result.tensor_buft_overrides = params_spec.tensor_buft_overrides;

+        // a draft pinned to a single device doesn't need the meta wrapper an inherited -sm tensor would give it
+        // (the device list is null-terminated, so a single device means size 2)
+        const size_t n_devs = std::count_if(params_spec.devices.begin(), params_spec.devices.end(),
+                [](ggml_backend_dev_t d) { return d != nullptr; });
+        if (n_devs == 1) {
+            result.split_mode = LLAMA_SPLIT_MODE_LAYER;
+        }
+
        if (params_spec.cpuparams.n_threads > 0) {
            result.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
            result.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
@@ -124,6 +124,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "HunYuanMoEV1ForCausalLM": "hunyuan",
    "HunYuanVLForConditionalGeneration": "hunyuan",
    "HYV3ForCausalLM": "hunyuan",
+    "HYV4ForCausalLM": "hy_v4",
    "IQuestCoderForCausalLM": "llama",
    "InternLM2ForCausalLM": "internlm",
    "InternLM3ForCausalLM": "internlm",
@@ -254,6 +255,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "SeedOssForCausalLM": "olmo",
    "SmallThinkerForCausalLM": "smallthinker",
    "SmolLM3ForCausalLM": "llama",
+    "Spark2_5ForCausalLM": "spark2_5",
    "SolarOpenForCausalLM": "glm",
    "StableLMEpochForCausalLM": "stablelm",
    "StableLmForCausalLM": "stablelm",
@@ -130,7 +130,8 @@ class ModelBase:
                 sentence_transformers_dense_modules: bool = False,
                 target_model_dir: Path | None = None,
                 fuse_gate_up_exps: bool = False,
-                 fp8_as_q8: bool = False):
+                 fp8_as_q8: bool = False,
+                 fuse_qkv: bool = False):
        if type(self) is ModelBase or \
                type(self) is TextModel or \
                type(self) is MmprojModel:
@@ -153,6 +154,15 @@ class ModelBase:
        self.fuse_gate_up_exps = fuse_gate_up_exps
        self._gate_exp_buffer: dict[int, Tensor] = {}
        self._up_exp_buffer: dict[int, Tensor] = {}
+        self.fuse_qkv = fuse_qkv
+        self._q_buffer: dict[int, Tensor] = {}
+        self._k_buffer: dict[int, Tensor] = {}
+        self._v_buffer: dict[int, Tensor] = {}
+        self._q_bias_buffer: dict[int, Tensor] = {}
+        self._k_bias_buffer: dict[int, Tensor] = {}
+        self._v_bias_buffer: dict[int, Tensor] = {}
+        self._fusable_qkv_weight_layers: set[int] = set()
+        self._fusable_qkv_bias_layers: set[int] = set()
        self.hparams = ModelBase.load_hparams(self.dir_model, self.is_mistral_format) if hparams is None else hparams
        self.model_tensors = self.index_tensors(remote_hf_model_id=remote_hf_model_id)
        self.metadata_override = metadata_override
@@ -617,6 +627,43 @@ class ModelBase:
            raise ValueError(f"Can not map tensor {name!r}")
        return new_name

+    def prepare_qkv_fusion(self) -> None:
+        self._fusable_qkv_weight_layers.clear()
+        self._fusable_qkv_bias_layers.clear()
+        if not self.fuse_qkv or gguf.MODEL_TENSOR.ATTN_QKV not in gguf.MODEL_TENSORS[self.model_arch]:
+            return
+
+        qkv_types = {
+            gguf.MODEL_TENSOR.ATTN_Q,
+            gguf.MODEL_TENSOR.ATTN_K,
+            gguf.MODEL_TENSOR.ATTN_V,
+        }
+        weights: dict[int, set[gguf.MODEL_TENSOR]] = {}
+        biases: dict[int, set[gguf.MODEL_TENSOR]] = {}
+
+        for name in self.model_tensors:
+            mapped = self.tensor_map.get_type_and_name(name, try_suffixes=(".weight", ".bias"))
+            if mapped is None:
+                continue
+            tensor_type, new_name = mapped
+            if tensor_type not in qkv_types:
+                continue
+
+            bid = next((int(part) for part in new_name.split(".") if part.isdecimal()), None)
+            if bid is None:
+                continue
+            if new_name.endswith(".weight"):
+                weights.setdefault(bid, set()).add(tensor_type)
+            elif new_name.endswith(".bias"):
+                biases.setdefault(bid, set()).add(tensor_type)
+
+        for bid, weight_types in weights.items():
+            bias_types = biases.get(bid, set())
+            if weight_types == qkv_types and (not bias_types or bias_types == qkv_types):
+                self._fusable_qkv_weight_layers.add(bid)
+                if bias_types:
+                    self._fusable_qkv_bias_layers.add(bid)
+
    def set_gguf_parameters(self):
        raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")

@@ -645,6 +692,40 @@ class ModelBase:
               self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
                return []

+        # Handle Q/K/V tensor fusion if enabled
+        qkv_bid = next((int(part) for part in new_name.split(".") if part.isdecimal()), None) if self.fuse_qkv else None
+        if qkv_bid is not None:
+            is_bias = new_name.endswith('.bias')
+            suffix = '.bias' if is_bias else '.weight'
+            fusable_layers = self._fusable_qkv_bias_layers if is_bias else self._fusable_qkv_weight_layers
+            if qkv_bid not in fusable_layers:
+                return [(new_name, data_torch)]
+
+            buf_q = self._q_bias_buffer if is_bias else self._q_buffer
+            buf_k = self._k_bias_buffer if is_bias else self._k_buffer
+            buf_v = self._v_bias_buffer if is_bias else self._v_buffer
+
+            if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_Q, qkv_bid, suffix):
+                buf_q[qkv_bid] = data_torch
+            elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_K, qkv_bid, suffix):
+                buf_k[qkv_bid] = data_torch
+            elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_V, qkv_bid, suffix):
+                buf_v[qkv_bid] = data_torch
+
+            if qkv_bid in buf_q and qkv_bid in buf_k and qkv_bid in buf_v:
+                q_data = buf_q.pop(qkv_bid)
+                k_data = buf_k.pop(qkv_bid)
+                v_data = buf_v.pop(qkv_bid)
+                fused_data = torch.cat([q_data, k_data, v_data], dim=0)
+                fused_name = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, qkv_bid, suffix=suffix)
+                logger.info(f"Fused Q, K, V {suffix[1:]} into QKV for layer {qkv_bid}")
+                return [(fused_name, fused_data)]
+
+            if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_Q, qkv_bid, suffix) or \
+               self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_K, qkv_bid, suffix) or \
+               self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_V, qkv_bid, suffix):
+                return []
+
        return [(new_name, data_torch)]

    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
@@ -899,6 +980,8 @@ class ModelBase:

        self.dequant_model()

+        self.prepare_qkv_fusion()
+
        # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
        if self.tensor_map.mapping:
            max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -1027,6 +1110,13 @@ class ModelBase:

                self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype)

+        qkv_buffers = (
+            self._q_buffer, self._k_buffer, self._v_buffer,
+            self._q_bias_buffer, self._k_bias_buffer, self._v_bias_buffer,
+        )
+        if any(qkv_buffers):
+            raise ValueError("QKV fusion did not consume all buffered tensors")
+
    def set_type(self):
        self.gguf_writer.add_type(gguf.GGUFType.MODEL)

@@ -1507,6 +1597,9 @@ class TextModel(ModelBase):
        if chkhsh == "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6":
            # ref: https://huggingface.co/tencent/Hunyuan-4B-Instruct
            res = "hunyuan-dense"
+        if chkhsh == "e6ddf9c6686791c12d698d34c31ab9be1fea9af5a3d9a6909783ab382198ae1c":
+            # ref: https://huggingface.co/tencent/Hy4-preview
+            res = "hy_v4"
        if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
            # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
            res = "falcon-h1"
@@ -1540,6 +1633,9 @@ class TextModel(ModelBase):
        if chkhsh == "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7":
            # ref: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B
            res = "lfm2"
+        if chkhsh == "0a766d034107bc736a3f2dc4968fd62e54a3570f1454443e0c5a4cc6bd7941ed":
+            # ref: https://huggingface.co/XHToken/Spark-X2.5-1.7B
+            res = "spark2_5"
        if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
            # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
            res = "llama-bpe"
@@ -0,0 +1,244 @@
+from __future__ import annotations
+
+import re
+from typing import Iterable
+
+import torch
+
+from .base import ModelBase, gguf, logger
+from .deepseek import DeepseekV2Model
+
+
+def split_gate_up(weight: torch.Tensor, moe_intermediate_size: int):
+    """Split a fused stacked gate_up expert tensor into (gate, up).
+
+    weight: [n_expert, 2*moe_intermediate_size, hidden] (gate first, up second).
+    Returns (gate, up) each [n_expert, moe_intermediate_size, hidden].
+    """
+    assert weight.shape[1] == 2 * moe_intermediate_size, f"{weight.shape[1]} != 2*{moe_intermediate_size}"
+    gate = weight[:, :moe_intermediate_size, :].contiguous()
+    up = weight[:, moe_intermediate_size:, :].contiguous()
+    return gate, up
+
+
+@ModelBase.register("HYV4ForCausalLM")
+@ModelBase.example("tencent/Hy4-preview")
+class HYV4Model(DeepseekV2Model):
+    """HY_V4: DeepSeek-V3 style MLA + MoE with iHC, a gated MLA output and a learnable sink.
+
+    Reuses DeepseekV2Model for the vocab and the MLA metadata, but overrides the tensor mapping
+    because HY_V4 ships pre-stacked / fused experts plus extra iHC, gate and sink tensors. The
+    rope rows are mapped straight through (no permute) - the graph rotates consecutive pairs.
+
+    DSA is supported: indexer weights are exported for the layers marked "full" in indexer_types.
+    "shared" layers reuse the top-k of the last preceding full layer at inference time, so they
+    carry no indexer weights.
+
+    MTP (num_nextn_predict_layers) is dropped, so the GGUF cannot be used for speculative
+    decoding. The reference only runs the MTP layers while training or while speculating, so they
+    cannot change single-token logits.
+    """
+
+    model_arch = gguf.MODEL_ARCH.HY_V4
+
+    merge_expert = False
+
+    # tensors a "full" indexer layer must carry
+    INDEXER_SUFFIXES = frozenset({
+        "self_attn.indexer.wq_b.weight",
+        "self_attn.indexer.wk.weight",
+        "self_attn.indexer.k_norm.weight",
+        "self_attn.indexer.k_norm.bias",
+        "self_attn.indexer.weights_proj.weight",
+    })
+
+    @classmethod
+    def filter_tensors(cls, item):
+        # drop MTP here, not in modify_tensors, so the weights are never read
+        if item[0].startswith("model.mtp_layers."):
+            return None
+        return super().filter_tensors(item)
+
+    def _check_indexer_hparams(self):
+        for key in ("index_n_heads", "index_head_dim", "index_topk"):
+            if key not in self.hparams:
+                raise ValueError(f"HY_V4 has DSA layers but no {key}")
+
+    def indexer_is_full(self) -> list[bool] | None:
+        """Per-layer indexer ownership, or None when the checkpoint has no DSA.
+
+        indexer_types entries are "full" (owns an indexer) or "shared" (reuses the preceding
+        full layer's top-k). Missing indexer_types with sparse layers means every sparse layer
+        owns one.
+        """
+        hparams = self.hparams
+        n_layer = hparams["num_hidden_layers"]
+        indexer_types = hparams.get("indexer_types")
+
+        # the reference drives DSA off indexer_types alone; layer_types is only a fallback for
+        # checkpoints predating it (it was renamed to deepseek_sparse_attention upstream)
+        if indexer_types is None:
+            layer_types = hparams.get("layer_types") or []
+            sparse = {"sparse_attention", "deepseek_sparse_attention"}
+            if not any(t in sparse for t in layer_types):
+                return None
+            if len(layer_types) < n_layer:
+                raise ValueError(f"HY_V4 layer_types has {len(layer_types)} entries, need {n_layer}")
+            self._check_indexer_hparams()
+            return [t in sparse for t in layer_types[:n_layer]]
+
+        self._check_indexer_hparams()
+
+        if len(indexer_types) < n_layer:
+            raise ValueError(f"HY_V4 indexer_types has {len(indexer_types)} entries, need {n_layer}")
+        unknown = {t for t in indexer_types[:n_layer]} - {"full", "shared"}
+        if unknown:
+            raise ValueError(f"HY_V4 unknown indexer_types values: {sorted(unknown)}")
+        is_full = [t == "full" for t in indexer_types[:n_layer]]
+        if is_full and not is_full[0]:
+            raise ValueError("HY_V4 layer 0 must be indexer_types 'full' (nothing precedes it to share)")
+        return is_full
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        # HY4 has n_group == topk_group == 1 (no group routing). Drop the keys so the base does
+        # not emit expert_group_count/used; llama.cpp then takes the ungrouped MoE path.
+        if hparams.get("n_group") == 1 and hparams.get("topk_group") == 1:
+            hparams.pop("n_group", None)
+            hparams.pop("topk_group", None)
+
+        # HY_V4 config expresses dense/sparse layers via mlp_layer_types, but DeepseekV2Model
+        # needs first_k_dense_replace. Derive it as the contiguous leading "dense" block
+        # (the real config.json also carries first_k_dense_replace; prefer it when present,
+        # but assert the two agree so a mismatch fails loudly).
+        mlp_types = hparams.get("mlp_layer_types")
+        explicit = hparams.get("first_k_dense_replace")
+        derived = None
+        if mlp_types is not None:
+            lead = 0
+            for t in mlp_types:
+                if t == "dense":
+                    lead += 1
+                else:
+                    break
+            if any(t == "dense" for t in mlp_types[lead:]):
+                raise NotImplementedError("HY_V4 converter expects a contiguous leading dense block")
+            derived = lead
+        if explicit is not None and derived is not None and explicit != derived:
+            raise ValueError(
+                f"HY_V4 first_k_dense_replace ({explicit}) disagrees with mlp_layer_types "
+                f"leading-dense count ({derived})"
+            )
+        if explicit is None:
+            if derived is None:
+                raise ValueError("HY_V4 needs first_k_dense_replace or mlp_layer_types to place dense layers")
+            hparams["first_k_dense_replace"] = derived
+
+        # reuse DeepseekV2 MLA + MoE metadata (forces num_key_value_heads=1, writes q/kv lora,
+        # key/value lengths, expert counts, weights scale/norm, rope dims, etc.)
+        super().set_gguf_parameters()
+
+        # HY4 uses DeepSeek-V3 sigmoid routing with e_score_correction_bias. The config has no
+        # scoring_func key, so the base does not write a gating func; set it explicitly.
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+        # routed-expert SwiGLU logits clamp (only routed experts; shared/dense are not clamped,
+        # so swiglu_clamp_shexp is intentionally not written). 0.0 disables the clamp.
+        swiglu_limit = float(hparams.get("swiglu_limit", 0.0) or 0.0)
+        if swiglu_limit > 0.0:
+            self.gguf_writer.add_swiglu_clamp_exp([swiglu_limit] * self.block_count)
+
+        # iHC (independent Hyper-Connections)
+        self.gguf_writer.add_hyper_connection_count(hparams["hc_mult"])
+        self.gguf_writer.add_hyper_connection_epsilon(hparams["hc_eps"])
+        self.gguf_writer.add_hyper_connection_magnitude(hparams["hc_magnitude"])
+
+        # is_full is written explicitly; the graph must not infer it from tensor presence
+        is_full = self.indexer_is_full()
+        if is_full is not None:
+            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(is_full)
+            logger.info(
+                "HY_V4 DSA: %d/%d layers own an indexer (top_k=%d, n_heads=%d, head_dim=%d)",
+                sum(is_full), len(is_full), hparams["index_topk"],
+                hparams["index_n_heads"], hparams["index_head_dim"],
+            )
+
+        if hparams.get("num_nextn_predict_layers", 0):
+            logger.warning(
+                "HY_V4: dropping %d MTP (nextn) layer(s) - the reference runs them only under "
+                "training / speculative decoding. This GGUF cannot be used for speculative decoding.",
+                hparams["num_nextn_predict_layers"],
+            )
+
+    def prepare_tensors(self):
+        # Hy4-preview for some reason has num_key_value_heads equal to 8, so override it here
+        # without this conversion/deepseek.py fails on assert
+        self.hparams["num_key_value_heads"] = self.hparams["num_attention_heads"]
+
+        # validate before the base materializes tensors, so a mismatch fails early
+        is_full = self.indexer_is_full()
+        if is_full is not None:
+            present: dict[int, set[str]] = {}
+            for name in self.model_tensors:
+                m = re.match(r"model\.layers\.(\d+)\.(self_attn\.indexer\..+)$", name)
+                if m:
+                    present.setdefault(int(m.group(1)), set()).add(m.group(2))
+            for il, expect_full in enumerate(is_full):
+                seen = present.get(il, set())
+                if expect_full and seen != self.INDEXER_SUFFIXES:
+                    raise ValueError(
+                        f"HY_V4 layer {il} is indexer_types 'full' but is missing indexer tensors: "
+                        f"{sorted(self.INDEXER_SUFFIXES - seen)}"
+                    )
+                if not expect_full and seen:
+                    raise ValueError(
+                        f"HY_V4 layer {il} is indexer_types 'shared' but carries indexer tensors: "
+                        f"{sorted(seen)}"
+                    )
+
+        super().prepare_tensors()
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # iHC mixing matrices are 2D .weight tensors that the reference keeps in fp32
+        # (_keep_in_fp32_modules_strict). 1D tensors (hc_base/scale, attn_sinks,
+        # e_score_correction_bias) and the router (FFN_GATE_INP) are already forced F32 by the
+        # base rules. Force the HC *_fn matrices here.
+        if new_name.endswith(("hc_attn_fn.weight", "hc_ffn_fn.weight", "output_hc_fn.weight")):
+            return gguf.GGMLQuantizationType.F32
+        # indexer k_norm is fp32 in the reference; the base rules already cover
+        # *_norm.weight and INDEXER_PROJ, but not this bias
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.INDEXER_K_NORM, bid, suffix=".bias"):
+            return gguf.GGMLQuantizationType.F32
+        # enable_lm_head_fp32: mirror the reference fp32 LM-head matmul by keeping output F32.
+        if new_name == "output.weight" and self.hparams.get("enable_lm_head_fp32", False):
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
+        hparams = self.hparams
+        moe_inter = hparams["moe_intermediate_size"]
+
+        tn = self.format_tensor_name
+
+        # fused stacked experts: split gate_up into gate/up
+        if name.endswith("mlp.experts.gate_up_proj"):
+            gate, up = split_gate_up(data_torch, moe_inter)
+            yield from super().modify_tensors(gate, tn(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid)
+            yield from super().modify_tensors(up,   tn(gguf.MODEL_TENSOR.FFN_UP_EXP,   bid), bid)
+            return
+
+        # add .weight suffixes
+        if name.endswith("mlp.experts.down_proj") or name.endswith(".self_attn.learnable_sink_param"):
+            name += ".weight"
+
+        if re.search(r"\.hc_head\.hc_head_(?:fn|base|scale)$", name):
+            name += ".weight"
+
+        if re.search(r"\.hc_(?:attn|mlp)_layer\.hc_pre\.hc_(?:fn|base|scale)$", name):
+            name += ".weight"
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -25,7 +25,7 @@ class MiniMaxText01Model(TextModel):
        # they get in the way of the token sampling process and must be suppressed

        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        tokenizer_vocab_size = tokenizer.vocab_size
+        tokenizer_vocab_size = tokenizer.vocab_size  # ty: ignore[unresolved-attribute]

        with open(self.dir_model / "model.safetensors.index.json", "r", encoding="utf-8") as f:
            weight_map = json.load(f)["weight_map"]
@@ -37,7 +37,7 @@ class MuseGlimmerModel(TextModel):

        from transformers import AutoTokenizer
        tok = AutoTokenizer.from_pretrained(self.dir_model)
-        eot_id = tok.convert_tokens_to_ids("<|eot|>")
+        eot_id = tok.convert_tokens_to_ids("<|eot|>")  # ty: ignore[unresolved-attribute]
        if isinstance(eot_id, int) and eot_id >= 0:
            self.gguf_writer.add_eot_token_id(eot_id)

@@ -379,6 +379,13 @@ class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
        self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
        self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
        self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
+        if (layer_types := self.hparams.get("layer_types")) is not None:
+            n_layer = self.hparams["num_hidden_layers"]
+            if len(layer_types) != n_layer:
+                raise ValueError(f"layer_types has {len(layer_types)} entries, expected num_hidden_layers ({n_layer})")
+            recurrent = [t == "linear_attention" for t in layer_types]
+            recurrent += [False] * (self.block_count - n_layer)
+            self.gguf_writer.add_recurrent_layers(recurrent)
        self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4))
        if (rope_dim := self.hparams.get("head_dim")) is None:
            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
@@ -0,0 +1,65 @@
+from __future__ import annotations
+
+from collections.abc import Iterable
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, TextModel, gguf
+
+
+@ModelBase.register("Spark2_5ForCausalLM")
+@ModelBase.example("XHToken/Spark-X2.5-1.7B")
+class Spark2_5Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.SPARK2_5
+
+    def set_gguf_parameters(self) -> None:
+        super().set_gguf_parameters()
+
+        hparams = self.hparams
+        layer_types = hparams["layer_types"]
+        if len(layer_types) != self.block_count:
+            raise ValueError(
+                f"Spark2_5 layer_types length {len(layer_types)} != num_hidden_layers {self.block_count}"
+            )
+        if any(layer_type not in ("sliding_attention", "full_attention") for layer_type in layer_types):
+            raise ValueError(f"Spark2_5 has unsupported layer_types: {layer_types}")
+        if hparams.get("gate_attn_act_mode") != "sigmoid" or hparams.get("headwise_attn_output_gate") is not True:
+            raise ValueError("Spark2_5 conversion requires head-wise sigmoid attention gates")
+        if hparams.get("hidden_act") != "gelu":
+            raise ValueError(f"Spark2_5 conversion requires GELU, got {hparams.get('hidden_act')!r}")
+
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern(
+            [layer_type == "sliding_attention" for layer_type in layer_types]
+        )
+
+        head_dim = hparams["head_dim"]
+        full_rope = self.rope_parameters["full_attention"]
+        swa_rope = self.rope_parameters["sliding_attention"]
+        self.gguf_writer.add_rope_dimension_count(
+            int(head_dim * float(full_rope["partial_rotary_factor"]))
+        )
+        self.gguf_writer.add_rope_dimension_count_swa(
+            int(head_dim * float(swa_rope["partial_rotary_factor"]))
+        )
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith(".self_attn.q_k_v_proj.weight"):
+            if bid is None:
+                raise ValueError(f"Spark2_5 fused QKV tensor has no block id: {name}")
+            yield self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, bid), data_torch
+            return
+
+        if name.endswith(".self_attn.g_proj.weight"):
+            if bid is None:
+                raise ValueError(f"Spark2_5 attention gate tensor has no block id: {name}")
+            expected = self.hparams["num_attention_heads"]
+            if data_torch.shape[0] != expected:
+                raise ValueError(
+                    f"Spark2_5 layer {bid} attention gate width {data_torch.shape[0]} != head count {expected}"
+                )
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -157,6 +157,10 @@ def parse_args() -> argparse.Namespace:
        help="Store tensors dequantized from FP8 as Q8_0 instead of BF16/F16.",
    )

+    parser.add_argument(
+        "--fuse-qkv", action="store_true",
+        help="Fuse separate Q, K, V weight tensors into a single QKV tensor.",
+    )
    parser.add_argument(
        "--target-model-dir", type=str, default=None,
        help=(
@@ -290,6 +294,7 @@ def main() -> None:
                                     target_model_dir=Path(args.target_model_dir) if args.target_model_dir else None,
                                     fuse_gate_up_exps=args.fuse_gate_up_exps,
                                     fp8_as_q8=args.fp8_as_q8,
+                                     fuse_qkv=args.fuse_qkv,
                                     )

        if args.vocab_only:
@@ -176,6 +176,7 @@ pre_computed_hashes = [
    {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
    {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
    {"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},
+    {"name": "hy_v4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hy4-preview", "chkhsh": "e6ddf9c6686791c12d698d34c31ab9be1fea9af5a3d9a6909783ab382198ae1c"},
    # falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
    {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
@@ -190,6 +191,7 @@ pre_computed_hashes = [
    {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/evilfreelancer/ruGPT3XL", "chkhsh": "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4"},
    # lfm2 variants
    {"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B", "chkhsh": "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7"},
+    {"name": "spark2_5", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/XHToken/Spark-X2.5-1.7B", "chkhsh": "0a766d034107bc736a3f2dc4968fd62e54a3570f1454443e0c5a4cc6bd7941ed"},
 ]


@@ -514,6 +514,7 @@ The following templates have active tests in `tests/test-chat.cpp`:
 | Mistral Small 3.2 | JSON_NATIVE | `[TOOL_CALLS]func[ARGS]{...}` with call ID |
 | Devstral | JSON_NATIVE | `[TOOL_CALLS]func[ARGS]{...}` without call ID |
 | StepFun 3.5 Flash | TAG_WITH_TAGGED | `<function=X><parameter=Y>` format |
+| Spark2.5 | TAG_WITH_TAGGED | `<tool_call>name<arg_key>...<arg_value>...` format |

 ## Adding Support for New Templates

@@ -805,6 +805,8 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
 | GGML_SYCL_ENABLE_MKL_FA | 1 (default) or 0 | Enable oneMKL GEMM flash attention for XMX-accelerated prompt processing with quantized KV cache. Automatically activates during prefill (prompt processing) when all conditions are met: (1) flash-attn enabled (`-fa` or `--flash-attn on`), (2) KV cache quantized (`--cache-type-k q8_0 --cache-type-v q8_0` or other `*_0/*_1` types), (3) batch size ≥ 1024 (`--batch-size 1024`), (4) prompt length ≥ 1024 tokens. Set to 0 to force the TILE kernel for A/B testing. Example minimum command: `llama-cli -m model.gguf -fa -ngl 99 --cache-type-k q8_0 --cache-type-v q8_0 --batch-size 1024 -p "your prompt"` |
 | GGML_SYCL_MKL_FA_DEBUG | 0 (default) or 1 | Enable per-call diagnostic logging for MKL flash attention: GEMM/softmax timings, interleaved-head detection, and buffer memory usage. |
+| GGML_SYCL_MEMTRACE | 0 (default), 1, 2 | Enable record and output memory allocation diagnostics. Requires `-lv 4`. <br>0 -  Disable<br>1 - Basic memory info, including current and peak allocations, as well allocations from other sources, around 50 lines per model load.<br>2 - More verbose, logging around 900 specific allocations and deallocations. |
+| GGML_SYCL_MEMTRACE_STEP | 64 (default) or positive integer | With GGML_SYCL_MEMTRACE=1, the minimum growth in memory usage to trigger another log record. |
 | GGML_SYCL_MKL_FA_DIAG | 0 (default) or 1 | Enable output fingerprinting for MKL flash attention. Dumps the first 64 float output values for the first 6 FA calls with n_kv ≥ 1024, labeled with kernel type (MKL/TILE/VEC) for cross-kernel comparison. |
 | GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute. Unsupported types and layouts fall back to the standalone op kernels. See `ggml_sycl_can_fuse()`. |
 | GGML_SYCL_ENABLE_ESIMD | 0 or 1 (default)| Enable ESIMD kernels when available. |
@@ -1177,7 +1177,7 @@ def create_dynamic_model_from_function(func: Callable[..., Any]):
        dynamic_fields[param.name] = (
            param.annotation if param.annotation != inspect.Parameter.empty else str, default_value)
    # Creating the dynamic model
-    dynamic_model = create_model(f"{getattr(func, '__name__')}", **dynamic_fields)
+    dynamic_model = create_model(f"{getattr(func, '__name__')}", **dynamic_fields)  # ty: ignore[no-matching-overload]

    for name, param_doc in param_docs:
        dynamic_model.model_fields[name].description = param_doc.description
@@ -2,8 +2,9 @@
 #include <cstdio>

 int main(void) {
-    printf("[test-cmake] version: %s, build: %d (%s)\n",
+    printf("[test-cmake] llama.cpp version: %s, build: %d (%s)\n",
           llama_version(), LLAMA_BUILD_NUMBER, LLAMA_BUILD_COMMIT);
+    printf("[test-cmake] ggml version: %s, commit: %s\n", ggml_version(), ggml_commit());
    printf("[test-cmake] Initializing backend...\n");
    llama_backend_init();
    printf("[test-cmake] Backend initialized.\n");
@@ -128,7 +128,7 @@
          }:
          {
            # For standardised reproducible formatting with `nix fmt`
-            formatter = pkgs.nixfmt-rfc-style;
+            formatter = pkgs.nixfmt;

            # Unlike `.#packages`, legacyPackages may contain values of
            # arbitrary types (including nested attrsets) and may even throw
@@ -156,7 +156,7 @@
                windows = config.legacyPackages.llamaPackagesWindows.llama-cpp;
                python-scripts = config.legacyPackages.llamaPackages.python-scripts;
              }
-              // lib.optionalAttrs pkgs.stdenv.isLinux {
+              // lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux {
                cuda = config.legacyPackages.llamaPackagesCuda.llama-cpp;

                mpi-cpu = config.packages.default.override { useMpi = true; };
@@ -433,10 +433,21 @@ extern "C" {
        GGML_TYPE_COUNT   = 43,
    };

-    // precision
+    // [TAG_GGML_PREC]
+    // this enum is used to declare the allowed numerical precision/data-types types that can be used during the compute of an op
+    // the declared types can be:
+    //  - result accumulation type
+    //  - source tensor data representation type
+    //  - etc.
+    // the precision parameters are stored as ggml_tensor.op_params to the respective ops
    enum ggml_prec {
-        GGML_PREC_DEFAULT =  0, // stored as ggml_tensor.op_params, 0 by default
-        GGML_PREC_F32     = 10,
+        GGML_PREC_UNDEFINED = 0,
+        GGML_PREC_DEFAULT   = 0,  // note: deprecated, use GGML_PREC_UNDEFINED
+        GGML_PREC_F32       = 10,
+        GGML_PREC_BF16      = 15,
+        GGML_PREC_F16       = 20,
+        GGML_PREC_Q8        = 30,
+        GGML_PREC_Q4        = 40,
    };

    // op hint
@@ -1429,6 +1440,42 @@ extern "C" {
            struct ggml_tensor  * b,
            float                 eps);

+    // [TAG_GGML_PREC]
+    // set the minimum required accumulator type for the implementation to use during the compute
+    // for example:
+    //  - GGML_PREC_F32  - requires accumulation of the results in F32
+    //  - GGML_PREC_BF16 - can accumulate the results in BF16, F32
+    //  - GGML_PREC_F16  - can accumulate the results in F16, F32
+    //  - GGML_PREC_Q8   - not allowed
+    //  - GGML_PREC_Q4   - not allowed
+    //
+    // return false on faliure
+    GGML_API bool ggml_prec_set_acc(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec);
+
+    // [TAG_GGML_PREC]
+    // set the smallest rank that the implementation can use to internally convert the src[idx] data to
+    // ranks in decreasing order:
+    //  - GGML_PREC_F32  - GGML_TYPE_F32
+    //  - GGML_PREC_BF16 - GGML_TYPE_BF16
+    //  - GGML_PREC_F16  - GGML_TYPE_F16,
+    //  - GGML_PREC_Q8   - GGML_TYPE_Q8_0, GGML_TYPE_Q8_1, GGML_TYPE_Q8_K, etc.
+    //  - GGML_PREC_Q4   - GGML_TYPE_Q4_0, GGML_TYPE_Q4_1, GGML_TYPE_Q4_K, GGML_TYPE_NVFP4, GGML_TYPE_MXFP4, etc.
+    //
+    // for example:
+    //   - ggml_prec_set_src(a, GGML_PREC_Q8, 1):
+    //     - allows the implementation to quantize F32, BF16, F16 data of src[1] down to GGML_TYPE_Q8_0
+    //     - cannot quantize it down to GGML_TYPE_Q4_0 or GGML_TYPE_NVFP4
+    //   - ggml_prec_set_src(a, GGML_PREC_Q4, 1):
+    //     - allows the implementation to quantize F32, BF16, F16 data of src[1] down to 4-bit datatypes such as GGML_TYPE_Q4_K, GGML_TYPE_NVFP4 etc.
+    //
+    // return false on faliure
+    GGML_API bool ggml_prec_set_src(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec,
+            int                  idx);
+
    // A: k columns, n rows => [ne03, ne02, n, k]
    // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
    // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
@@ -1439,9 +1486,10 @@ extern "C" {

    // change the precision of a matrix multiplication
    // set to GGML_PREC_F32 for higher precision (useful for phi-2)
-    GGML_API void ggml_mul_mat_set_prec(
+    GGML_DEPRECATED(GGML_API void ggml_mul_mat_set_prec(
            struct ggml_tensor * a,
-            enum ggml_prec       prec);
+            enum ggml_prec       prec),
+        "use ggml_prec_set_acc() instead");

    // change the hint of a matrix multiplication
    GGML_API void ggml_mul_mat_set_hint(
@@ -2446,9 +2494,10 @@ extern "C" {
            float                 max_bias,
            float                 logit_softcap);

-    GGML_API void ggml_flash_attn_ext_set_prec(
+    GGML_DEPRECATED(GGML_API void ggml_flash_attn_ext_set_prec(
            struct ggml_tensor * a,
-            enum ggml_prec       prec);
+            enum ggml_prec       prec),
+        "use ggml_prec_set_acc() instead");

    GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
            const struct ggml_tensor * a);
@@ -849,7 +849,7 @@ static void ggml_backend_sched_split_inputs_grow(struct ggml_backend_sched_split
    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
    if (split->inputs_capacity > 0) {
        new_cap = 2*split->inputs_capacity;
-        GGML_LOG_WARN("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
+        GGML_LOG_DEBUG("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
    }
    auto * pnew = (struct ggml_tensor **) realloc((void *) split->inputs, new_cap * sizeof(struct ggml_tensor *));
    if (pnew == NULL) {
@@ -864,7 +864,7 @@ static void ggml_backend_sched_graph_inputs_grow(ggml_backend_sched_t sched) {
    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
    if (sched->graph_inputs_capacity > 0) {
        new_cap = 2*sched->graph_inputs_capacity;
-        GGML_LOG_WARN("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
+        GGML_LOG_DEBUG("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
    }
    auto * pnew = (struct ggml_tensor **) realloc((void *) sched->graph_inputs, new_cap * sizeof(struct ggml_tensor *));
    if (pnew == NULL) {
@@ -1338,17 +1338,6 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                            break;
                        }
                    }
-                    // check if the split has too many inputs
-                    // FIXME: count the number of inputs instead of only checking when full
-                    if (split->n_inputs >= split->inputs_capacity) {
-                        const size_t id = hash_id(src);
-                        int src_backend_id = sched->hv_tensor_backend_ids[id];
-                        bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
-                        if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
-                            need_new_split = true;
-                            break;
-                        }
-                    }
                }
            }

@@ -78,7 +78,7 @@ struct ggml_compute_params {
 #if defined(__ARM_NEON)

 // ref: https://github.com/ggml-org/llama.cpp/pull/5404
-#ifdef _MSC_VER
+#if defined(_MSC_VER) && !defined(__clang__)
 #define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
 #else
 #define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
@@ -69,6 +69,8 @@
 #define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 0x803)  // Tonga, Fiji, Polaris, minimum for fast fp16
 #define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 0x900)  // Vega56/64, minimum for fp16 dual issue
 #define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 0x906)  // MI50/Radeon VII, minimum for dp4a
+#define GGML_CUDA_CC_GFX909     (GGML_CUDA_CC_OFFSET_AMD + 0x909)  // GCN APU
+#define GGML_CUDA_CC_GFX90C     (GGML_CUDA_CC_OFFSET_AMD + 0x90c)  // GCN APU
 #define GGML_CUDA_CC_CDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
 #define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x90a)  // MI210 (gfx90a), minimum acc register renaming
 #define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300
@@ -89,12 +91,13 @@
 #define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_RDNA3(cc)   (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc))
 #define GGML_CUDA_CC_IS_RDNA4(cc)   (cc >= GGML_CUDA_CC_RDNA4)
-#define GGML_CUDA_CC_IS_GCN(cc)     (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)
-#define GGML_CUDA_CC_IS_CDNA(cc)    (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
-#define GGML_CUDA_CC_IS_CDNA1(cc)   (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
-#define GGML_CUDA_CC_IS_CDNA2(cc)   (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
-#define GGML_CUDA_CC_IS_CDNA3(cc)   (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4)
-#define GGML_CUDA_CC_IS_CDNA4(cc)   (cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_GCN_APU(cc) ((cc) == GGML_CUDA_CC_GFX909 || (cc) == GGML_CUDA_CC_GFX90C)
+#define GGML_CUDA_CC_IS_GCN(cc)     ((cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1) || GGML_CUDA_CC_IS_GCN_APU(cc))
+#define GGML_CUDA_CC_IS_CDNA(cc)    (!GGML_CUDA_CC_IS_GCN_APU(cc) && cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_CDNA1(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
+#define GGML_CUDA_CC_IS_CDNA2(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
+#define GGML_CUDA_CC_IS_CDNA3(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4)
+#define GGML_CUDA_CC_IS_CDNA4(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1)

 // Moore Threads
 #define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons
@@ -121,6 +124,12 @@
 #    define GGML_CUDA_USE_PDL
 #endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && (CUDART_VERSION >= 12030 || (!(defined(_MSC_VER) && !defined(__clang__)) && CUDART_VERSION >= 11080))

+static __device__ __forceinline__ void ggml_cuda_syncwarp() {
+#ifndef GGML_USE_HIP
+    __syncwarp();
+#endif // GGML_USE_HIP
+}
+
 static __device__ __forceinline__ void ggml_cuda_pdl_sync() {
 #if defined(GGML_CUDA_USE_PDL) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_HOPPER
    cudaGridDependencySynchronize();
@@ -970,6 +979,7 @@ template<>
 struct ggml_cuda_type_traits<GGML_TYPE_F16> {
    static constexpr int qk = 1;
    static constexpr int qr = 1;
+    static constexpr int bs = sizeof(ggml_half);
 };

 template<>
@@ -977,6 +987,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q1_0> {
    static constexpr int qk = QK1_0;
    static constexpr int qr = QR1_0;
    static constexpr int qi = QI1_0;
+    static constexpr int bs = sizeof(block_q1_0);
 };

 template<>
@@ -984,6 +995,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q2_0> {
    static constexpr int qk = QK2_0;
    static constexpr int qr = QR2_0;
    static constexpr int qi = QI2_0;
+    static constexpr int bs = sizeof(block_q2_0);
 };

 template<>
@@ -991,6 +1003,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
    static constexpr int qk = QK4_0;
    static constexpr int qr = QR4_0;
    static constexpr int qi = QI4_0;
+    static constexpr int bs = sizeof(block_q4_0);
 };

 template<>
@@ -998,6 +1011,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q4_1> {
    static constexpr int qk = QK4_1;
    static constexpr int qr = QR4_1;
    static constexpr int qi = QI4_1;
+    static constexpr int bs = sizeof(block_q4_1);
 };

 template<>
@@ -1005,6 +1019,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q5_0> {
    static constexpr int qk = QK5_0;
    static constexpr int qr = QR5_0;
    static constexpr int qi = QI5_0;
+    static constexpr int bs = sizeof(block_q5_0);
 };

 template<>
@@ -1012,6 +1027,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q5_1> {
    static constexpr int qk = QK5_1;
    static constexpr int qr = QR5_1;
    static constexpr int qi = QI5_1;
+    static constexpr int bs = sizeof(block_q5_1);
 };

 template<>
@@ -1019,6 +1035,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q8_0> {
    static constexpr int qk = QK8_0;
    static constexpr int qr = QR8_0;
    static constexpr int qi = QI8_0;
+    static constexpr int bs = sizeof(block_q8_0);
 };

 template<>
@@ -1026,6 +1043,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_MXFP4> {
    static constexpr int qk = QK_MXFP4;
    static constexpr int qr = QR_MXFP4;
    static constexpr int qi = QI_MXFP4;
+    static constexpr int bs = sizeof(block_mxfp4);
 };

 template<>
@@ -1033,6 +1051,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_NVFP4> {
    static constexpr int qk = QK_NVFP4;
    static constexpr int qr = QR_NVFP4;
    static constexpr int qi = QI_NVFP4;
+    static constexpr int bs = sizeof(block_nvfp4);
 };

 template<>
@@ -1040,6 +1059,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR2_K;
    static constexpr int qi = QI2_K;
+    static constexpr int bs = sizeof(block_q2_K);
 };

 template<>
@@ -1047,6 +1067,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q3_K> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR3_K;
    static constexpr int qi = QI3_K;
+    static constexpr int bs = sizeof(block_q3_K);
 };

 template<>
@@ -1054,6 +1075,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q4_K> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR4_K;
    static constexpr int qi = QI4_K;
+    static constexpr int bs = sizeof(block_q4_K);
 };

 template<>
@@ -1061,6 +1083,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q5_K> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR5_K;
    static constexpr int qi = QI5_K;
+    static constexpr int bs = sizeof(block_q5_K);
 };

 template<>
@@ -1068,6 +1091,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q6_K> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR6_K;
    static constexpr int qi = QI6_K;
+    static constexpr int bs = sizeof(block_q6_K);
 };

 template<>
@@ -1075,6 +1099,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XXS> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR2_XXS;
    static constexpr int qi = QI2_XXS;
+    static constexpr int bs = sizeof(block_iq2_xxs);
 };

 template<>
@@ -1082,6 +1107,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XS> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR2_XS;
    static constexpr int qi = QI2_XS;
+    static constexpr int bs = sizeof(block_iq2_xs);
 };

 template<>
@@ -1089,6 +1115,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ2_S> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR2_S;
    static constexpr int qi = QI2_S;
+    static constexpr int bs = sizeof(block_iq2_s);
 };

 template<>
@@ -1096,6 +1123,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ3_XXS> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR3_XXS;
    static constexpr int qi = QI3_XXS;
+    static constexpr int bs = sizeof(block_iq3_xxs);
 };

 template<>
@@ -1103,6 +1131,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ1_S> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR1_S;
    static constexpr int qi = QI1_S;
+    static constexpr int bs = sizeof(block_iq1_s);
 };

 template<>
@@ -1110,6 +1139,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ1_M> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR1_M;
    static constexpr int qi = QI1_M;
+    static constexpr int bs = sizeof(block_iq1_m);
 };

 template<>
@@ -1117,6 +1147,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ4_NL> {
    static constexpr int qk = QK4_NL;
    static constexpr int qr = QR4_NL;
    static constexpr int qi = QI4_NL;
+    static constexpr int bs = sizeof(block_iq4_nl);
 };

 template<>
@@ -1124,6 +1155,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ4_XS> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR4_XS;
    static constexpr int qi = QI4_XS;
+    static constexpr int bs = sizeof(block_iq4_xs);
 };

 template<>
@@ -1131,6 +1163,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ3_S> {
    static constexpr int qk = QK_K;
    static constexpr int qr = QR3_S;
    static constexpr int qi = QI3_S;
+    static constexpr int bs = sizeof(block_iq3_s);
 };

 //////////////////////
@@ -1545,77 +1545,77 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        }
    }

-    if (np > 1 && threadIdx.y % np == 0) {
-        // Combine the meta data for parallel warps via shared memory.
-        // Warps with threadIdx.y % np != 0 must NOT return early.
-        // All threads must return simultaneously to avoid race conditions with work on the next tile.
-
+    if (np > 1) {
        constexpr int nmeta = np*cols_per_warp >= warp_size ? np*cols_per_warp/warp_size : 1;

+        float KQ_cmn;
+        float KQ_cms[nmeta];
+        float KQ_crs;
+
        const int jc_meta = threadIdx.y*cols_per_warp + (np*cols_per_warp < warp_size ? threadIdx.x % (np*cols_per_warp) : threadIdx.x);
        float2 * const meta_ptr = ((float2 *) tile_Q) + jc_meta*(tile_stride/2) + nbatch_combine/2;
-        float2 meta[nmeta];
-#pragma unroll
-        for (int imeta = 0; imeta < nmeta; ++imeta) {
-            meta[imeta] = meta_ptr[imeta * warp_size * tile_stride/2];
-        }

-        float KQ_cmn = meta[0].x; // KQ combine max new, max between all parallel warps.
+        if (threadIdx.y % np == 0) {
+            // Combine the meta data for parallel warps via shared memory.
+            float2 meta[nmeta];
 #pragma unroll
-        for (int imeta = 1; imeta < nmeta; ++imeta) {
-            KQ_cmn = fmaxf(KQ_cmn, meta[imeta].x);
-        }
-#pragma unroll
-        for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
-            if (offset < warp_size) {
-                KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, warp_size));
+            for (int imeta = 0; imeta < nmeta; ++imeta) {
+                meta[imeta] = meta_ptr[imeta * warp_size * tile_stride/2];
            }
-        }

-        float KQ_cms[nmeta]; // KQ combine max scale per warp.
+            KQ_cmn = meta[0].x; // KQ combine max new, max between all parallel warps.
 #pragma unroll
-        for (int imeta = 0; imeta < nmeta; ++imeta) {
-            KQ_cms[imeta] = expf(meta[imeta].x - KQ_cmn);
-        }
+            for (int imeta = 1; imeta < nmeta; ++imeta) {
+                KQ_cmn = fmaxf(KQ_cmn, meta[imeta].x);
+            }
+#pragma unroll
+            for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
+                if (offset < warp_size) {
+                    KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, warp_size));
+                }
+            }

-        float KQ_crs = KQ_cms[0]*meta[0].y; // KQ combine rowsum, scaled sum of all parallel warps.
 #pragma unroll
-        for (int imeta = 1; imeta < nmeta; ++imeta) {
-            KQ_crs += KQ_cms[imeta]*meta[imeta].y;
-        }
+            for (int imeta = 0; imeta < nmeta; ++imeta) {
+                KQ_cms[imeta] = expf(meta[imeta].x - KQ_cmn);
+            }
+
+            KQ_crs = KQ_cms[0]*meta[0].y; // KQ combine rowsum, scaled sum of all parallel warps.
 #pragma unroll
-        for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
-            if (offset < warp_size) {
-                KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, warp_size);
+            for (int imeta = 1; imeta < nmeta; ++imeta) {
+                KQ_crs += KQ_cms[imeta]*meta[imeta].y;
+            }
+#pragma unroll
+            for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
+                if (offset < warp_size) {
+                    KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, warp_size);
+                }
            }
        }

        __syncthreads();

-        // Write back combined meta data:
+        if (threadIdx.y % np == 0) {
+            // Write back combined meta data:
 #pragma unroll
-        for (int imeta = 0; imeta < nmeta; ++imeta) {
-            if (np*cols_per_warp >= warp_size || threadIdx.x < np*cols_per_warp) {
-                // Combined KQ max scale + rowsum.
-                meta_ptr[imeta * warp_size * tile_stride/2] = make_float2(KQ_cms[imeta], KQ_crs);
+            for (int imeta = 0; imeta < nmeta; ++imeta) {
+                if (np*cols_per_warp >= warp_size || threadIdx.x < np*cols_per_warp) {
+                    // Combined KQ max scale + rowsum.
+                    meta_ptr[imeta * warp_size * tile_stride/2] = make_float2(KQ_cms[imeta], KQ_crs);
+                }
+            }
+
+            // Combined KQ max + rowsum.
+            static_assert(cols_per_warp <= warp_size);
+            if (needs_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
+                float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols;
+                dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
+            }
+            if (is_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
+                float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols;
+                dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
            }
        }
-
-        // Combined KQ max + rowsum.
-        static_assert(cols_per_warp <= warp_size);
-        if (needs_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
-            float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols;
-            dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
-        }
-        if (is_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
-            float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols;
-            dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
-        }
-    } else if (np > 1) {
-        // Warps with threadIdx.y % np == 0 execute a __syncthreads() in the if branch.
-        // Therefore, all other warps also need to execute a __syncthreads().
-        // Otherwise the points at which warps synchronize with each other would become misaligned.
-        __syncthreads();
    }

 #pragma unroll
@@ -317,9 +317,7 @@ static __global__ void flash_attn_ext_vec(
 #endif // V_DOT2_F32_F16_AVAILABLE
        }

-#ifndef GGML_USE_HIP
-        __syncwarp();
-#endif // GGML_USE_HIP
+        ggml_cuda_syncwarp();

 #pragma unroll
        for (int k0 = 0; k0 < WARP_SIZE; k0 += V_cols_per_iter) {
@@ -212,6 +212,7 @@ static int ggml_cuda_parse_id(char devName[]) {
    }
    archNum += archMajor * 0x100;
    archNum += archMinor;
+
    return archNum;
 }
 #endif // defined(GGML_USE_HIP)
@@ -303,11 +304,7 @@ static ggml_cuda_device_info ggml_cuda_init() {

        info.default_tensor_split[id] = total_vram;
        total_vram += device_vram;
-#if defined(GGML_USE_HIP)
-        info.devices[id].integrated = prop.integrated;
-#else
        info.devices[id].integrated = false; // Temporarily disabled due to issues with corrupted output (e.g. #15034)
-#endif
        info.devices[id].nsm        = prop.multiProcessorCount;
        info.devices[id].smpb       = prop.sharedMemPerBlock;
        info.devices[id].warp_size  = prop.warpSize;
@@ -143,6 +143,7 @@ static __global__ void mul_mat_f(
            if (threadIdx.x == 0) {
                slot_map[j] = -1;
            }
+            ggml_cuda_syncwarp();

            if (col_base + j >= ncols_dst_total) {
                continue;
@@ -171,10 +172,12 @@ static __global__ void mul_mat_f(
        tile_A A[ntA][warp_size / tile_A::J];
 #pragma unroll
        for (int itA = 0; itA < ntA; ++itA) {
+            ggml_cuda_syncwarp();
 #pragma unroll
            for (int i = 0; i < tile_A::I; ++i) {
                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
            }
+            ggml_cuda_syncwarp();
 #pragma unroll
            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
@@ -183,6 +186,7 @@ static __global__ void mul_mat_f(

 #pragma unroll
        for (int itB = 0; itB < ntB; ++itB) {
+            ggml_cuda_syncwarp();
            if constexpr (std::is_same_v<T, float>) {
 #pragma unroll
                for (int j0 = 0; j0 < tile_B::I; ++j0) {
@@ -212,6 +216,7 @@ static __global__ void mul_mat_f(
            } else {
                static_assert(std::is_same_v<T, void>, "unsupported type");
            }
+            ggml_cuda_syncwarp();
 #pragma unroll
            for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
                tile_B B;
@@ -229,6 +234,8 @@ static __global__ void mul_mat_f(

    if (nwarps > 1) {
        __syncthreads();
+    } else {
+        ggml_cuda_syncwarp();
    }
 #pragma unroll
    for (int itB = 0; itB < ntB; ++itB) {
@@ -245,6 +252,8 @@ static __global__ void mul_mat_f(

    if (nwarps > 1) {
        __syncthreads();
+    } else {
+        ggml_cuda_syncwarp();
    }

 #pragma unroll
@@ -382,10 +391,12 @@ static __global__ void mul_mat_f_ids(
        tile_A A[ntA][warp_size / tile_A::J];
 #pragma unroll
        for (int itA = 0; itA < ntA; ++itA) {
+            ggml_cuda_syncwarp();
 #pragma unroll
            for (int i = 0; i < tile_A::I; ++i) {
                tile_xy[i*tile_k_padded + threadIdx.x] = x[(itA*tile_A::I + i)*stride_row  + col];
            }
+            ggml_cuda_syncwarp();
 #pragma unroll
            for (int k0 = 0; k0 < warp_size; k0 += tile_A::J) {
                load_ldmatrix(A[itA][k0/tile_A::J], tile_xy + k0, tile_k_padded);
@@ -419,6 +430,7 @@ static __global__ void mul_mat_f_ids(
            int next_buf = 1;
 #pragma unroll
            for (int itB = 0; itB < ntB; ++itB) {
+                ggml_cuda_syncwarp();
 #pragma unroll
                for (int j0 = 0; j0 < tile_B::I; ++j0) {
                    tile_xy[j0*tile_k_padded + threadIdx.x] = vals_buf[curr_buf][j0];
@@ -428,6 +440,7 @@ static __global__ void mul_mat_f_ids(
                    gather_tile(itB + 1, vals_buf[next_buf]);
                }

+                ggml_cuda_syncwarp();
 #pragma unroll
                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
                    tile_B B;
@@ -472,6 +485,7 @@ static __global__ void mul_mat_f_ids(
            int next_buf = 1;
 #pragma unroll
            for (int itB = 0; itB < ntB; ++itB) {
+                ggml_cuda_syncwarp();
 #pragma unroll
                for (int j0 = 0; j0 < tile_B::I; ++j0) {
                    const float2 tmp = vals_buf[curr_buf][j0];
@@ -482,6 +496,7 @@ static __global__ void mul_mat_f_ids(
                    gather_tile(itB + 1, vals_buf[next_buf]);
                }

+                ggml_cuda_syncwarp();
 #pragma unroll
                for (int k0 = 0; k0 < warp_size; k0 += tile_B::J) {
                    tile_B B;
@@ -507,6 +522,8 @@ static __global__ void mul_mat_f_ids(

    if (nwarps > 1) {
        __syncthreads();
+    } else {
+        ggml_cuda_syncwarp();
    }
 #pragma unroll
    for (int itB = 0; itB < ntB; ++itB) {
@@ -523,6 +540,8 @@ static __global__ void mul_mat_f_ids(

    if (nwarps > 1) {
        __syncthreads();
+    } else {
+        ggml_cuda_syncwarp();
    }

 #pragma unroll
@@ -101,6 +101,7 @@ static __global__ void mm_ids_helper(
        }
    }
    nex_prev = warp_reduce_sum<warp_size>(nex_prev);
+    ggml_cuda_syncwarp();

    for (int itc = threadIdx.x; itc < it_compact; itc += warp_size) {
        const mm_ids_helper_store store_it = store[itc];
@@ -375,10 +375,10 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
        return true;
    }

-    // gfx900 (Vega 10) lacks native dp4a, loses to dequant + hipBLAS
+    // gfx900 (Vega 10), gfx909, and gfx90c lack native dp4a, losing to dequant + hipBLAS
    // for dense matrices; keep MMQ only for MoE, where the
    // hipBLAS path is much slower.
-    if (cc == GGML_CUDA_CC_VEGA) {
+    if (cc == GGML_CUDA_CC_VEGA || GGML_CUDA_CC_IS_GCN_APU(cc)) {
        return n_experts > 0;
    }

@@ -6,6 +6,35 @@
 #include <cstdint>
 #include <type_traits>

+// only enabled on DGX Spark, where it is a gain on every type below. On the higher-bandwidth parts the kernel
+// has little exposed latency left to hide and the extra requests cost more than they save.
+// For perf data, see https://github.com/ggml-org/llama.cpp/pull/26705#issuecomment-5569335031
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == GGML_CUDA_CC_DGX_SPARK
+// returns true only for those quants that benefit from prefetch and false otherwise
+static constexpr __host__ __device__ bool mmvq_should_prefetch(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_MXFP4:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static __device__ __forceinline__ void mmvq_prefetch_l2(const void * p) {
+    asm volatile("prefetch.global.L2 [%0];" :: "l"(p));
+}
+#endif
+
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);

 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
@@ -298,9 +327,6 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
                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;
        }
@@ -310,8 +336,9 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
            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_Q5_K:
+                return ne11 <= 6;
            case GGML_TYPE_Q6_K:
                return ne11 <= 7;
            default:
@@ -675,6 +702,26 @@ static __global__ void mul_mat_vec_q(
        // x block quant index when casting the quants to int
        const int kqs = vdr * (tid % (qi/vdr));

+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == GGML_CUDA_CC_DGX_SPARK
+        // start the next iterations' weight loads early
+        if constexpr (mmvq_should_prefetch(type)) {
+            constexpr int pf_dist = 2; // loop iterations, not blocks
+            const int kbx_pf = kbx + pf_dist*blocks_per_iter;
+            if (kbx_pf < blocks_per_row_x) {
+#pragma unroll
+                for (int i = 0; i < rows_per_cuda_block; ++i) {
+                    const size_t off = (size_t)(kbx_offset + i*stride_row_x + kbx_pf) * ggml_cuda_type_traits<type>::bs;
+                    mmvq_prefetch_l2((const char *) vx + off);
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            mmvq_prefetch_l2((const char *) vgate + off);
+                        }
+                    }
+                }
+            }
+        }
+#endif
+
 #pragma unroll
        for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
@@ -936,16 +936,20 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
    v[0] = q4[0];
    v[1] = q4[4];

+    // branchless so nvcc can hoist this out of the ncols_dst loop
    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
+    const int j  = bq8_offset/2;
+    const int jm = j & 1;
+
+    const uint32_t s0 = scales[jm + 0];
+    const uint32_t s2 = scales[jm + 2];
+    const uint32_t s4 = scales[jm + 4];
+
+    const uint32_t hi = (uint32_t) -(int32_t) (j >= 2);
+
    uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
+    aux[0] = (uint16_t) (((s0 & 0x3f3f) & ~hi) | ((((s4 >> 0) & 0x0f0f) | ((s0 & 0xc0c0) >> 2)) & hi));
+    aux[1] = (uint16_t) (((s2 & 0x3f3f) & ~hi) | ((((s4 >> 4) & 0x0f0f) | ((s2 & 0xc0c0) >> 2)) & hi));
    const uint8_t * sc = (const uint8_t *)aux;
    const uint8_t * m  = sc + 2;

@@ -981,16 +985,21 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
    vh[0] = qh[0] >> bq8_offset;
    vh[1] = qh[4] >> bq8_offset;

+    // same as q4_K
    const uint16_t * scales = (const uint16_t *)bq5_K->scales;
+    const int j  = bq8_offset/2;
+    const int jm = j & 1;
+
+    const uint32_t s0 = scales[jm + 0];
+    const uint32_t s2 = scales[jm + 2];
+    const uint32_t s4 = scales[jm + 4];
+
+    const uint32_t hi = (uint32_t) -(int32_t) (j >= 2);
+
    uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
+    aux[0] = (uint16_t) (((s0 & 0x3f3f) & ~hi) | ((((s4 >> 0) & 0x0f0f) | ((s0 & 0xc0c0) >> 2)) & hi));
+    aux[1] = (uint16_t) (((s2 & 0x3f3f) & ~hi) | ((((s4 >> 4) & 0x0f0f) | ((s2 & 0xc0c0) >> 2)) & hi));
+
    const uint8_t * sc = (const uint8_t *)aux;
    const uint8_t * m  = sc + 2;

@@ -176,9 +176,9 @@

 #define __CUDA_ARCH__ 1300

-#if defined(__gfx900__) || defined(__gfx906__)
+#if defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)
 #define GCN5
-#endif // defined(__gfx900__) || defined(__gfx906__)
+#endif // defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)

 #if defined(__gfx803__)
 #define GCN4
@@ -4979,6 +4979,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
        case GGML_OP_CONCAT:          return HTP_OP_CONCAT;
        case GGML_OP_SCALE:           return HTP_OP_SCALE;
        case GGML_OP_CLAMP:           return HTP_OP_CLAMP;
+        case GGML_OP_LEAKY_RELU:      return HTP_OP_LEAKY_RELU;
        case GGML_OP_SQR:             return HTP_OP_SQR;
        case GGML_OP_SQRT:            return HTP_OP_SQRT;
        case GGML_OP_LOG:             return HTP_OP_UNARY_LOG;
@@ -5006,6 +5007,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
                case GGML_UNARY_OP_SOFTPLUS:   return HTP_OP_UNARY_SOFTPLUS;
                case GGML_UNARY_OP_TANH:       return HTP_OP_UNARY_TANH;
                case GGML_UNARY_OP_ABS:        return HTP_OP_UNARY_ABS;
+                case GGML_UNARY_OP_RELU:       return HTP_OP_UNARY_RELU;
            default:
                break;
            }
@@ -5871,6 +5873,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
        case GGML_OP_RMS_NORM:
        case GGML_OP_SCALE:
        case GGML_OP_CLAMP:
+        case GGML_OP_LEAKY_RELU:
            supp = ggml_hexagon_supported_unary(sess, op);
            break;

@@ -5899,6 +5902,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
                case GGML_UNARY_OP_SILU:
                case GGML_UNARY_OP_GELU:
                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_RELU:
                    supp = ggml_hexagon_supported_unary(sess, op);
                    break;
                default:
@@ -65,6 +65,7 @@ enum htp_op_code {
    HTP_OP_UNARY_TANH,
    HTP_OP_UNARY_ABS,
    HTP_OP_UNARY_LOG,
+    HTP_OP_UNARY_RELU,
    HTP_OP_GLU_SWIGLU,
    HTP_OP_GLU_SWIGLU_OAI,
    HTP_OP_GLU_GEGLU,
@@ -93,6 +94,7 @@ enum htp_op_code {
    HTP_OP_NORM,
    HTP_OP_CONCAT,
    HTP_OP_CLAMP,
+    HTP_OP_LEAKY_RELU,
    HTP_OP_IM2COL,
    HTP_OP_FENCE,
    HTP_OP_ALLREDUCE,
@@ -308,6 +308,46 @@ static inline void hvx_min_scalar_f32(uint8_t * restrict dst, const uint8_t * re
    }
 }

+// MAX Scalar variants
+
+#define HVX_OP_MAX_SCALAR(v) Q6_Vsf_vmax_VsfVsf(val_vec, v)
+
+static inline void hvx_max_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
+    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(float), hvx_vec_store_a, HVX_OP_MAX_SCALAR);
+}
+
+static inline void hvx_max_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
+    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
+    assert((unsigned long) dst % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(float), hvx_vec_store_a, HVX_OP_MAX_SCALAR);
+}
+
+static inline void hvx_max_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
+    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(float), hvx_vec_store_u, HVX_OP_MAX_SCALAR);
+}
+
+static inline void hvx_max_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float val, uint32_t n) {
+    const HVX_Vector val_vec = hvx_vec_splat_f32(val);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(float), hvx_vec_store_u, HVX_OP_MAX_SCALAR);
+}
+
+static inline void hvx_max_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float val, const int num_elems) {
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
+        hvx_max_scalar_f32_aa(dst, src, val, num_elems);
+    } else if (hex_is_aligned((void *) dst, 128)) {
+        hvx_max_scalar_f32_au(dst, src, val, num_elems);
+    } else if (hex_is_aligned((void *) src, 128)) {
+        hvx_max_scalar_f32_ua(dst, src, val, num_elems);
+    } else {
+        hvx_max_scalar_f32_uu(dst, src, val, num_elems);
+    }
+}
+
 // CLAMP Scalar variants

 #define HVX_OP_CLAMP_SCALAR(v) \
@@ -406,6 +446,53 @@ static inline void hvx_clamp_scalar_f16(uint8_t * restrict dst, const uint8_t *
    }
 }

+#define HVX_OP_LEAKY_RELU_SCALAR(v)                                 \
+    ({                                                              \
+        HVX_VectorPred pred_neg = Q6_Q_vcmp_gt_VsfVsf(zero_vec, v); \
+        HVX_Vector     scaled   = HVX_OP_MUL_F32(v, ns_vec);        \
+        Q6_V_vmux_QVV(pred_neg, scaled, v);                         \
+    })
+
+static inline void hvx_leaky_relu_scalar_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const float ns, uint32_t n) {
+    const HVX_Vector zero_vec = hvx_vec_splat_f32(0.0f);
+    const HVX_Vector ns_vec   = hvx_vec_splat_f32(ns);
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(float), hvx_vec_store_a, HVX_OP_LEAKY_RELU_SCALAR);
+}
+
+static inline void hvx_leaky_relu_scalar_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const float ns, uint32_t n) {
+    const HVX_Vector zero_vec = hvx_vec_splat_f32(0.0f);
+    const HVX_Vector ns_vec   = hvx_vec_splat_f32(ns);
+    assert((unsigned long) dst % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(float), hvx_vec_store_a, HVX_OP_LEAKY_RELU_SCALAR);
+}
+
+static inline void hvx_leaky_relu_scalar_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const float ns, uint32_t n) {
+    const HVX_Vector zero_vec = hvx_vec_splat_f32(0.0f);
+    const HVX_Vector ns_vec   = hvx_vec_splat_f32(ns);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(float), hvx_vec_store_u, HVX_OP_LEAKY_RELU_SCALAR);
+}
+
+static inline void hvx_leaky_relu_scalar_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const float ns, uint32_t n) {
+    const HVX_Vector zero_vec = hvx_vec_splat_f32(0.0f);
+    const HVX_Vector ns_vec   = hvx_vec_splat_f32(ns);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(float), hvx_vec_store_u, HVX_OP_LEAKY_RELU_SCALAR);
+}
+
+static inline void hvx_leaky_relu_scalar_f32(uint8_t * restrict dst, const uint8_t * restrict src, const float ns, const int num_elems) {
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
+        hvx_leaky_relu_scalar_f32_aa(dst, src, ns, num_elems);
+    } else if (hex_is_aligned((void *) dst, 128)) {
+        hvx_leaky_relu_scalar_f32_au(dst, src, ns, num_elems);
+    } else if (hex_is_aligned((void *) src, 128)) {
+        hvx_leaky_relu_scalar_f32_ua(dst, src, ns, num_elems);
+    } else {
+        hvx_leaky_relu_scalar_f32_uu(dst, src, ns, num_elems);
+    }
+}
+
 //
 // Abs
 //
@@ -627,8 +714,10 @@ static inline void hvx_sqr_f16(uint8_t * restrict dst, const uint8_t * restrict
 #undef HVX_OP_MUL_SCALAR_F16
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
+#undef HVX_OP_MAX_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
 #undef HVX_OP_CLAMP_SCALAR_F16
+#undef HVX_OP_LEAKY_RELU_SCALAR
 #undef DEFINE_HVX_BINARY_OP_VARIANTS
 #undef HVX_BINARY_DISPATCHER
 #undef UNUSED
@@ -771,6 +771,7 @@ static int execute_op(struct htp_ops_context * octx) {
        case HTP_OP_RMS_NORM_MUL:
        case HTP_OP_SCALE:
        case HTP_OP_CLAMP:
+        case HTP_OP_LEAKY_RELU:
        case HTP_OP_SQR:
        case HTP_OP_SQRT:
        case HTP_OP_UNARY_SOFTPLUS:
@@ -782,6 +783,7 @@ static int execute_op(struct htp_ops_context * octx) {
        case HTP_OP_UNARY_TANH:
        case HTP_OP_UNARY_ABS:
        case HTP_OP_UNARY_LOG:
+        case HTP_OP_UNARY_RELU:
        case HTP_OP_L2_NORM:
            return op_unary(octx);

@@ -156,6 +156,22 @@ static void clamp_f32(const float * restrict src,
    }
 }

+static void leaky_relu_f32(const float * restrict src,
+                           float * restrict dst,
+                           const uint32_t num_rows,
+                           const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float negative_slope = 0.f;
+    memcpy(&negative_slope, &op_params[0], sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_leaky_relu_scalar_f32(dst_local, src_local, negative_slope, ne0);
+    }
+}
+
 static void rms_norm_f32(const float * restrict src,
                         float * restrict dst,
                         const uint32_t num_rows,
@@ -597,6 +613,20 @@ static void abs_f32(const float * restrict src,
    }
 }

+static void relu_f32(const float * restrict src,
+                     float * restrict dst,
+                     const uint32_t num_rows,
+                     const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_max_scalar_f32(dst_local, src_local, 0.0f, ne0);
+    }
+}
+
 static void log_f32(const float * restrict src,
                    float * restrict dst,
                    const uint32_t num_rows,
@@ -774,6 +804,7 @@ DEFINE_UNARY_TASK(rms_norm,       false, false, rms_norm_f32(src0_vtcm, dst_vtcm
 DEFINE_UNARY_TASK(rms_norm_mul,   true,  false, rms_norm_mul_f32(src0_vtcm, uctx->broadcast_weight ? (const float *) src1_vtcm_data : src1_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(scale,          false, false, scale_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(clamp,          false, false, clamp_f32(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK(leaky_relu,     false, false, leaky_relu_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(sqr,            false, false, sqr_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(sqrt,           false, false, sqrt_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(unary_neg,      false, false, neg_f32(src0_vtcm, dst_vtcm, block_size, uctx))
@@ -785,6 +816,7 @@ DEFINE_UNARY_TASK(unary_softplus, false, false, softplus_f32(src0_vtcm, dst_vtcm
 DEFINE_UNARY_TASK(unary_tanh,     false, false, tanh_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(unary_abs,      false, false, abs_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(unary_log,      false, false, log_f32(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK(unary_relu,     false, false, relu_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(l2_norm,        false, false, l2_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(tri,            false, true,  tri_f32(src0_vtcm, dst_vtcm, block_size, ir, uctx))

@@ -937,6 +969,12 @@ static inline void tile_clamp_f32(uint8_t * dst_vtcm, const uint8_t * src_vtcm,
    hvx_clamp_scalar_f32(dst_vtcm, src_vtcm, min, max, tw);
 }

+static inline void tile_leaky_relu_f32(uint8_t * dst_vtcm, const uint8_t * src_vtcm, uint32_t tw, const int32_t * op_params) {
+    float negative_slope = 0.f;
+    memcpy(&negative_slope, &op_params[0], sizeof(float));
+    hvx_leaky_relu_scalar_f32(dst_vtcm, src_vtcm, negative_slope, tw);
+}
+
 static inline void tile_unary_softplus_f32(uint8_t * dst_vtcm, const uint8_t * src_vtcm, uint32_t tw) {
    const float * restrict sf = (const float *) src_vtcm;
    float * restrict df       = (float *) dst_vtcm;
@@ -1035,6 +1073,7 @@ static inline void tri_apply_tile_f32(const uint8_t * restrict src, uint8_t * re

 DEFINE_UNARY_TILED_TASK(scale,          false, tile_scale_f32(dst_vtcm, src_vtcm, tw, op_params))
 DEFINE_UNARY_TILED_TASK(clamp,          false, tile_clamp_f32(dst_vtcm, src_vtcm, tw, op_params))
+DEFINE_UNARY_TILED_TASK(leaky_relu,     false, tile_leaky_relu_f32(dst_vtcm, src_vtcm, tw, op_params))
 DEFINE_UNARY_TILED_TASK(sqr,            false, hvx_sqr_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(sqrt,           false, hvx_sqrt_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(unary_neg,      false, hvx_scale_f32_aa(dst_vtcm, src_vtcm, tw, -1.0f))
@@ -1046,6 +1085,7 @@ DEFINE_UNARY_TILED_TASK(unary_softplus, false, tile_unary_softplus_f32(dst_vtcm,
 DEFINE_UNARY_TILED_TASK(unary_tanh,     false, hvx_tanh_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(unary_abs,      false, hvx_abs_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(unary_log,      false, hvx_log_f32_aa(dst_vtcm, src_vtcm, tw))
+DEFINE_UNARY_TILED_TASK(unary_relu,     false, hvx_max_scalar_f32(dst_vtcm, src_vtcm, 0.0f, tw))
 DEFINE_UNARY_TILED_TASK(tri,            true,  tri_apply_tile_f32(src_vtcm, dst_vtcm, tw, col, i01, ne0, tri_ttype))

 static int execute_op_unary(struct htp_ops_context * octx) {
@@ -1064,6 +1104,7 @@ static int execute_op_unary(struct htp_ops_context * octx) {
        case HTP_OP_RMS_NORM_MUL:    op_type = "rmsnorm-mul-f32";                            break;
        case HTP_OP_SCALE:           op_type = is_f16 ? "scale-f16"    : "scale-f32";        break;
        case HTP_OP_CLAMP:           op_type = is_f16 ? "clamp-f16"    : "clamp-f32";        break;
+        case HTP_OP_LEAKY_RELU:      op_type = "leaky-relu-f32";                             break;
        case HTP_OP_SQR:             op_type = is_f16 ? "sqr-f16"      : "sqr-f32";          break;
        case HTP_OP_SQRT:            op_type = is_f16 ? "sqrt-f16"     : "sqrt-f32";         break;
        case HTP_OP_UNARY_NEG:       op_type = "neg-f32";                                    break;
@@ -1075,9 +1116,9 @@ static int execute_op_unary(struct htp_ops_context * octx) {
        case HTP_OP_UNARY_TANH:      op_type = "tanh-f32";                                   break;
        case HTP_OP_UNARY_ABS:       op_type = is_f16 ? "abs-f16"      : "abs-f32";          break;
        case HTP_OP_UNARY_LOG:       op_type = is_f16 ? "log-f16"      : "log-f32";          break;
+        case HTP_OP_UNARY_RELU:      op_type = "relu-f32";                                   break;
        case HTP_OP_L2_NORM:         op_type = is_f16 ? "l2norm-f16"   : "l2norm-f32";       break;
        case HTP_OP_TRI:             op_type = "tri-f32";                                    break;
-
        default:
            FARF(ERROR, "Unsupported unary Op %u\n", octx->op);
            return HTP_STATUS_NO_SUPPORT;
@@ -1190,6 +1231,7 @@ static int execute_op_unary(struct htp_ops_context * octx) {
            switch (octx->op) {
                case HTP_OP_SCALE:           task_func = unary_task_f32_tiled_scale;          break;
                case HTP_OP_CLAMP:           task_func = unary_task_f32_tiled_clamp;          break;
+                case HTP_OP_LEAKY_RELU:      task_func = unary_task_f32_tiled_leaky_relu;     break;
                case HTP_OP_SQR:             task_func = unary_task_f32_tiled_sqr;            break;
                case HTP_OP_SQRT:            task_func = unary_task_f32_tiled_sqrt;           break;
                case HTP_OP_UNARY_NEG:       task_func = unary_task_f32_tiled_unary_neg;      break;
@@ -1201,6 +1243,7 @@ static int execute_op_unary(struct htp_ops_context * octx) {
                case HTP_OP_UNARY_TANH:      task_func = unary_task_f32_tiled_unary_tanh;     break;
                case HTP_OP_UNARY_ABS:       task_func = unary_task_f32_tiled_unary_abs;      break;
                case HTP_OP_UNARY_LOG:       task_func = unary_task_f32_tiled_unary_log;      break;
+                case HTP_OP_UNARY_RELU:      task_func = unary_task_f32_tiled_unary_relu;     break;
                case HTP_OP_TRI:             task_func = unary_task_f32_tiled_tri;            break;
                default:                     break;
            }
@@ -1224,6 +1267,7 @@ static int execute_op_unary(struct htp_ops_context * octx) {
                case HTP_OP_RMS_NORM_MUL:    task_func = unary_task_f32_rms_norm_mul;         break;
                case HTP_OP_SCALE:           task_func = unary_task_f32_scale;                break;
                case HTP_OP_CLAMP:           task_func = unary_task_f32_clamp;                break;
+                case HTP_OP_LEAKY_RELU:      task_func = unary_task_f32_leaky_relu;           break;
                case HTP_OP_SQR:             task_func = unary_task_f32_sqr;                  break;
                case HTP_OP_SQRT:            task_func = unary_task_f32_sqrt;                 break;
                case HTP_OP_UNARY_NEG:       task_func = unary_task_f32_unary_neg;            break;
@@ -1235,6 +1279,7 @@ static int execute_op_unary(struct htp_ops_context * octx) {
                case HTP_OP_UNARY_TANH:      task_func = unary_task_f32_unary_tanh;           break;
                case HTP_OP_UNARY_ABS:       task_func = unary_task_f32_unary_abs;            break;
                case HTP_OP_UNARY_LOG:       task_func = unary_task_f32_unary_log;            break;
+                case HTP_OP_UNARY_RELU:      task_func = unary_task_f32_unary_relu;           break;
                case HTP_OP_L2_NORM:         task_func = unary_task_f32_l2_norm;              break;
                case HTP_OP_TRI:             task_func = unary_task_f32_tri;                  break;
                default:                     break;
@@ -42,6 +42,7 @@ _Static_assert(sizeof(struct htp_unary_kernel_params) <= 128, "htp_unary_kernel_
 static inline bool htp_op_is_unary(uint32_t opcode) {
    switch (opcode) {
        case HTP_OP_CLAMP:
+        case HTP_OP_LEAKY_RELU:
        case HTP_OP_NORM:
        case HTP_OP_RMS_NORM:
        case HTP_OP_RMS_NORM_MUL:
@@ -57,6 +58,7 @@ static inline bool htp_op_is_unary(uint32_t opcode) {
        case HTP_OP_UNARY_TANH:
        case HTP_OP_UNARY_ABS:
        case HTP_OP_UNARY_LOG:
+        case HTP_OP_UNARY_RELU:
        case HTP_OP_L2_NORM:
        case HTP_OP_TRI:
            return true;
@@ -160,6 +160,18 @@ static float ggml_get_op_params_f32(const struct ggml_tensor * tensor, uint32_t
    return ((const float *)(tensor->op_params))[i];
 }

+// [TAG_GGML_PREC]
+// - GGML_OP_MUL_MAT
+//   0 - acc
+//   1 - hint
+//   2 - src0 precision
+//   3 - src1 precision
+//
+// - GGML_OP_MUL_MAT_ID
+//   0 - acc
+//   1 - hint
+//   2 - src0 precision
+//   3 - src1 precision
 static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
    assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
    ((int32_t *)(tensor->op_params))[i] = value;
@@ -111,6 +111,7 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(dev);
        if (queue == nil) {
            GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
+            free(res);
            return NULL;
        }

@@ -839,6 +839,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta

    const char * suffix = "";

+    bool split = false;
+
    // use custom matrix x vector kernel
    switch (tsrc0) {
        case GGML_TYPE_F32:
@@ -942,6 +944,13 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
                nsg = N_SG_IQ3_XXS;
                nr0 = N_R0_IQ3_XXS;
                smem = 256*4+128;
+
+                // split the rows across threads when there are fewer than 32 chunks per row
+                const int nb32 = ne00/32;
+                if (nb32 < 32 && (32 % nb32) == 0) {
+                    nr0 = N_R0_IQ3_XXS_SPLIT;
+                    split = true;
+                }
            } break;
        case GGML_TYPE_IQ3_S:
            {
@@ -993,7 +1002,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
    const int16_t r3 = (int16_t) (ne13 / ne03);

    snprintf(base, 256, "kernel_mul_mv_%s_%s%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1), suffix);
-    snprintf(name, 256, "%s_nsg=%d_ne12=%d_r2=%d_r3=%d", base, nsg, ne12, r2, r3);
+    snprintf(name, 256, "%s_nsg=%d_ne12=%d_r2=%d_r3=%d_split=%d", base, nsg, ne12, r2, r3, split);

    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
    if (!res.pipeline) {
@@ -1003,6 +1012,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
        ggml_metal_cv_set_int16(cv, (int16_t) ne12, FC_MUL_MV + 2);
        ggml_metal_cv_set_int16(cv, r2,             FC_MUL_MV + 3);
        ggml_metal_cv_set_int16(cv, r3,             FC_MUL_MV + 4);
+        ggml_metal_cv_set_bool (cv, split,          FC_MUL_MV + 5);

        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);

@@ -1081,6 +1091,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id(ggml_m

    const char * suffix = "";

+    bool split = false;
+
        // use custom matrix x vector kernel
    switch (tsrc0) {
        case GGML_TYPE_F32:
@@ -1177,6 +1189,13 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id(ggml_m
                nsg = N_SG_IQ3_XXS;
                nr0 = N_R0_IQ3_XXS;
                smem = 256*4+128;
+
+                // split the rows across threads when there are fewer than 32 chunks per row
+                const int nb32 = ne00/32;
+                if (nb32 < 32 && (32 % nb32) == 0) {
+                    nr0 = N_R0_IQ3_XXS_SPLIT;
+                    split = true;
+                }
            } break;
        case GGML_TYPE_IQ3_S:
            {
@@ -1224,7 +1243,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id(ggml_m
    };

    snprintf(base, 256, "kernel_mul_mv_id_%s_%s%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1), suffix);
-    snprintf(name, 256, "%s_nsg=%d", base, nsg);
+    snprintf(name, 256, "%s_nsg=%d_split=%d", base, nsg, split);

    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
    if (!res.pipeline) {
@@ -1234,6 +1253,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id(ggml_m
        ggml_metal_cv_set_int16(cv, 1,   FC_MUL_MV + 2);
        ggml_metal_cv_set_int16(cv, 1,   FC_MUL_MV + 3);
        ggml_metal_cv_set_int16(cv, 1,   FC_MUL_MV + 4);
+        ggml_metal_cv_set_bool (cv, split, FC_MUL_MV + 5);

        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);

@@ -1486,7 +1486,9 @@ static bool ggml_metal_supports_mul_mat_op(
        const struct ggml_tensor * op,
        bool src0_f16_has_mv,
        bool mm_path) {
-    if (!has_simdgroup_reduction || op->src[0]->type == GGML_TYPE_NVFP4) {
+    if (!has_simdgroup_reduction ||
+        op->src[0]->type == GGML_TYPE_NVFP4 ||
+        op->src[0]->type == GGML_TYPE_TQ1_0) {
        return false;
    }

@@ -1887,7 +1889,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                };
            }
        case GGML_OP_GET_ROWS:
-            return op->src[0]->type != GGML_TYPE_NVFP4;
+            return op->src[0]->type != GGML_TYPE_NVFP4 &&
+                   op->src[0]->type != GGML_TYPE_TQ1_0;
        case GGML_OP_SET_ROWS:
            {
                if (op->src[0]->type == GGML_TYPE_F16) {
@@ -77,6 +77,7 @@

 #define N_R0_IQ3_XXS 4
 #define N_SG_IQ3_XXS 2
+#define N_R0_IQ3_XXS_SPLIT 8

 #define N_R0_IQ3_S 4
 #define N_SG_IQ3_S 2
@@ -1248,6 +1248,153 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 512, 512, 2, 0 }, { 4, 4 } },
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 512, 512, 2, 1 }, { 4, 1 } },
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 512, 512, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 32, 32, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 32, 32, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 64, 64, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 64, 64, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 96, 96, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 96, 96, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 32, 32, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 32, 32, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 64, 64, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 64, 64, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 64, 64, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 96, 96, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 128, 128, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 192, 192, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 192, 192, 2, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 32, 32, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 192, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 192, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 192, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 128, 2, 3 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 192, 128, 3, 3 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 256, 256, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 32, 32, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 192, 128, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q5_1, 576, 512, -1, 1 }, { 1, 4 } },
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 2, 4 } },
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, 1, 2 }, { 1, 4 } },
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_Q8_0, 32, 32, 2, 2 }, { 4, 4 } },
@@ -1525,6 +1672,178 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 1 }, { 4, 2 } },
    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 2 }, { 4, 2 } },
    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 64, 64, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 64, 64, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 192, 3, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 192, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 192, 3, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 320, 256, 3, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 576, 512, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_0, 576, 512, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 32, 32, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 64, 64, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 96, 96, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 96, 96, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 128, 128, 1, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 128, 128, 3, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 128, 128, 3, 2 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 128, 128, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 320, 256, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 320, 256, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 320, 256, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 320, 256, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 320, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 320, 256, 3, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 512, 512, 2, 0 }, { 4, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 512, 512, 2, 4 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 512, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 512, 512, 3, 2 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q4_1, 576, 512, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 128, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 128, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 128, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 128, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 256, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 320, 256, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 320, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 512, 512, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 576, 512, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 576, 512, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_0, 576, 512, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 128, 128, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 128, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 128, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 128, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 128, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 128, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 128, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 320, 256, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 320, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 576, 512, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 576, 512, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 576, 512, 3, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q5_1, 576, 512, 3, 3 }, { 1, 4 } },
    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 4, 4 } },
    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 1, 1 }, { 2, 4 } },
    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 2, 1 }, { 2, 4 } },
@@ -213,6 +213,7 @@ constant short FC_mul_mv_nxpsg [[function_constant(FC_MUL_MV + 1)]];
 constant short FC_mul_mv_ne12  [[function_constant(FC_MUL_MV + 2)]];
 constant short FC_mul_mv_r2    [[function_constant(FC_MUL_MV + 3)]];
 constant short FC_mul_mv_r3    [[function_constant(FC_MUL_MV + 4)]];
+constant bool  FC_mul_mv_split [[function_constant(FC_MUL_MV + 5)]];

 template<typename block_q_type, short NR0, typename args_t>
 void mul_vec_q_n_f32_impl(
@@ -2092,6 +2093,7 @@ kernel void kernel_mul_mv_iq2_xs_f32(
    kernel_mul_mv_iq2_xs_f32_impl<N_R0_IQ2_XS, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }

+// FC_mul_mv_split: for nb32 < 32 (nb32 divides 32), 32/nb32 threads share each chunk and each takes a slice of the rows
 template<int nr0, typename args_t>
 void kernel_mul_mv_iq3_xxs_f32_impl(
        args_t args,
@@ -2138,11 +2140,18 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
        threadgroup_barrier(mem_flags::mem_threadgroup);
    }

-    const int ix = tiisg;
+    const short ntx  = FC_mul_mv_split ? nb32 : 32;
+    const short nrep = 32 / ntx;
+
+    const short ix   = tiisg % ntx;
+    const short irep = tiisg / ntx;
+
+    const short row0 = (nr0 * irep      ) / nrep;
+    const short row1 = (nr0 * (irep + 1)) / nrep;

    device const float * y4 = y + 32 * ix;

-    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+    for (int ib32 = ix; ib32 < nb32; ib32 += ntx) {
        for (short i = 0; i < 32; ++i) {
            yl[i] = y4[i];
        }
@@ -2151,11 +2160,11 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
        const int ib  = ib32 % (QK_K / 32);

        device const block_iq3_xxs * xr = x + ibl;
-        device const uint8_t  * q3 = xr->qs + 8 * ib;
-        device const uint16_t * gas = (device const uint16_t *)(xr->qs + QK_K/4) + 2 * ib;
-        device const half * dh = &xr->d;
+        device const uint8_t  * q3 = xr->qs + 8 * ib + (uint64_t) row0*args.nb01;
+        device const uint16_t * gas = (device const uint16_t *)(xr->qs + QK_K/4) + 2 * ib + (uint64_t) row0*args.nb01/2;
+        device const half * dh = &xr->d + (uint64_t) row0*args.nb01/2;

-        for (short row = 0; row < nr0; row++) {
+        for (short row = row0; row < row1; row++) {
            const float db = dh[0];
            const uint32_t aux32 = gas[0] | (gas[1] << 16);
            const float d = db * (0.5f + (aux32 >> 28));
@@ -2177,7 +2186,7 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
            gas += args.nb01/2;
        }

-        y4 += 32 * 32;
+        y4 += 32 * ntx;
    }

    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
@@ -2190,6 +2199,23 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
    }
 }

+template<typename args_t>
+void kernel_mul_mv_iq3_xxs_f32_disp(
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+    if (FC_mul_mv_split) {
+        kernel_mul_mv_iq3_xxs_f32_impl<N_R0_IQ3_XXS_SPLIT, args_t>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
+    } else {
+        kernel_mul_mv_iq3_xxs_f32_impl<N_R0_IQ3_XXS, args_t>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
+    }
+}
+
 [[host_name("kernel_mul_mv_iq3_xxs_f32")]]
 kernel void kernel_mul_mv_iq3_xxs_f32(
        constant ggml_metal_kargs_mul_mv & args,
@@ -2201,7 +2227,7 @@ kernel void kernel_mul_mv_iq3_xxs_f32(
        ushort tiisg[[thread_index_in_simdgroup]],
        ushort sgitg[[simdgroup_index_in_threadgroup]]) {

-    kernel_mul_mv_iq3_xxs_f32_impl<N_R0_IQ3_XXS, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq3_xxs_f32_disp<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }

 template<int nr0, typename args_t>
@@ -3217,7 +3243,7 @@ template [[host_name("kernel_mul_mv_id_iq1_s_f32")]]   kernel kernel_mul_mv_id_t
 template [[host_name("kernel_mul_mv_id_iq1_m_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq1_m_f32_impl  <N_R0_IQ1_M>>>;
 template [[host_name("kernel_mul_mv_id_iq2_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xxs_f32_impl<N_R0_IQ2_XXS>>>;
 template [[host_name("kernel_mul_mv_id_iq2_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xs_f32_impl <N_R0_IQ2_XS>>>;
-template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_xxs_f32_impl<N_R0_IQ3_XXS>>>;
+template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_xxs_f32_disp<ggml_metal_kargs_mul_mv>>>;
 template [[host_name("kernel_mul_mv_id_iq3_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_s_f32_impl  <N_R0_IQ3_S>>>;
 template [[host_name("kernel_mul_mv_id_iq2_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl  <N_R0_IQ2_S>>>;
 template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl <N_R0_IQ4_NL>>>;
@@ -222,6 +222,7 @@ set(GGML_OPENCL_KERNELS
    exp
    expm1
    abs
+    unary_ext
    softplus
    pad
    repeat
@@ -238,7 +239,7 @@ set(GGML_OPENCL_KERNELS
 )

 if (GGML_OPENCL_USE_ADRENO_KERNELS)
-    list(APPEND GGML_OPENCL_KERNELS gemm_xmem_f16_f32_os8)
+    list(APPEND GGML_OPENCL_KERNELS gemm_xmem_f16_f32_os8 sdpa_xmem_f32_f16_os8)
 endif ()

 foreach (K ${GGML_OPENCL_KERNELS})
@@ -1,56 +1,66 @@
-kernel void kernel_concat_f32(
-    global  const char * src0,
-    ulong                offset0,
-    global  const char * src1,
-    ulong                offset1,
-    global        char * dst,
-    ulong                offsetd,
-    int             ne00,
-    int             ne01,
-    int             ne02,
-    int             ne03,
-    ulong           nb00,
-    ulong           nb01,
-    ulong           nb02,
-    ulong           nb03,
-    ulong           nb10,
-    ulong           nb11,
-    ulong           nb12,
-    ulong           nb13,
-    int             ne0,
-    ulong           nb0,
-    ulong           nb1,
-    ulong           nb2,
-    ulong           nb3,
-    int             dim
-) {
-    src0 = src0 + offset0;
-    src1 = src1 + offset1;
-    dst  = dst  + offsetd;
+// concat is a pure copy, so the kernels are keyed by element byte size
+// (1/2/4/8) rather than logical type, matching the CUDA backend.

-    const int i3 = get_group_id(2);
-    const int i2 = get_group_id(1);
-    const int i1 = get_group_id(0);
-
-    int o[4] = {0, 0, 0, 0};
-    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
-
-    global const float * x;
-
-    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
-        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
-        } else {
-            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
-        }
-
-        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-
-        *y = *x;
-    }
+#define KERNEL_CONCAT(SUFFIX, T)                                                     \
+kernel void kernel_concat_##SUFFIX(                                                  \
+    global  const char * src0,                                                       \
+    ulong                offset0,                                                    \
+    global  const char * src1,                                                       \
+    ulong                offset1,                                                    \
+    global        char * dst,                                                        \
+    ulong                offsetd,                                                    \
+    int             ne00,                                                            \
+    int             ne01,                                                            \
+    int             ne02,                                                            \
+    int             ne03,                                                            \
+    ulong           nb00,                                                            \
+    ulong           nb01,                                                            \
+    ulong           nb02,                                                            \
+    ulong           nb03,                                                            \
+    ulong           nb10,                                                            \
+    ulong           nb11,                                                            \
+    ulong           nb12,                                                            \
+    ulong           nb13,                                                            \
+    int             ne0,                                                             \
+    ulong           nb0,                                                             \
+    ulong           nb1,                                                             \
+    ulong           nb2,                                                             \
+    ulong           nb3,                                                             \
+    int             dim                                                              \
+) {                                                                                  \
+    src0 = src0 + offset0;                                                           \
+    src1 = src1 + offset1;                                                           \
+    dst  = dst  + offsetd;                                                           \
+                                                                                     \
+    const int i3 = get_group_id(2);                                                  \
+    const int i2 = get_group_id(1);                                                  \
+    const int i1 = get_group_id(0);                                                  \
+                                                                                     \
+    int o[4] = {0, 0, 0, 0};                                                         \
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));         \
+                                                                                     \
+    global const T * x;                                                              \
+                                                                                     \
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {              \
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {                      \
+            x = (global const T *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00); \
+        } else {                                                                     \
+            x = (global const T *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10); \
+        }                                                                            \
+                                                                                     \
+        global T * y = (global T *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);        \
+                                                                                     \
+        *y = *x;                                                                     \
+    }                                                                                \
 }

-kernel void kernel_concat_f32_pack(
+KERNEL_CONCAT(b1, char)
+KERNEL_CONCAT(b2, short)
+KERNEL_CONCAT(b4, int)
+KERNEL_CONCAT(b8, long)
+
+// packed variant for the common dim==0, small-ne0 case (4-byte elements only).
+kernel void kernel_concat_b4_pack(
    global  const char * src0,
    ulong                offset0,
    global  const char * src1,
@@ -104,14 +114,14 @@ kernel void kernel_concat_f32_pack(
    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));

    for (int i0 = lane; i0 < ne0; i0 += tpr) {
-        global const float * x;
+        global const int * x;
        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+            x = (global const int *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
        } else {
-            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+            x = (global const int *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
        }

-        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+        global int * y = (global int *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);

        *y = *x;
    }
@@ -48,8 +48,8 @@ kernel void kernel_conv_2d(
    uint Cout, uint Cin, uint N,
    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
-    uint nb01, uint nb02, uint nb03,
-    uint nb11, uint nb12, uint nb13,
+    uint nb00, uint nb01, uint nb02, uint nb03,
+    uint nb10, uint nb11, uint nb12, uint nb13,
    uint nb1, uint nb2, uint nb3
 ) {
    global T_FLOAT* knl_data = (global T_FLOAT*) ((global char*)p_knl + off_knl);
@@ -95,7 +95,7 @@ kernel void kernel_conv_2d(
                const uint Cin_idx = crs_g / (KW*KH);
                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
-                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                const uint knl_idx = KW_idx*nb00 + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
            } else {
                Ash[k_l * BS_CRS + crs_l] = (T_FLOAT)0.0f;
@@ -123,7 +123,7 @@ kernel void kernel_conv_2d(
                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);

                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
-                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            const uint src_idx = W_idx * nb10 + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
                            ((T_FLOAT*)&val)[v] = src_data[src_idx];
                        }
                    }
@@ -39,8 +39,8 @@ kernel void kernel_conv_2d(
    uint Cout, uint Cin, uint N,
    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
-    uint nb01, uint nb02, uint nb03,
-    uint nb11, uint nb12, uint nb13,
+    uint nb00, uint nb01, uint nb02, uint nb03,
+    uint nb10, uint nb11, uint nb12, uint nb13,
    uint nb1, uint nb2, uint nb3
 ) {
    global half* knl_data = (global half*) ((global char*)p_knl + off_knl);
@@ -86,7 +86,7 @@ kernel void kernel_conv_2d(
                const uint Cin_idx = crs_g / (KW*KH);
                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
-                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                const uint knl_idx = KW_idx*nb00 + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
            } else {
                Ash[k_l * BS_CRS + crs_l] = (half)0.0f;
@@ -114,7 +114,7 @@ kernel void kernel_conv_2d(
                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);

                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
-                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            const uint src_idx = W_idx * nb10 + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
                            ((float*)&val)[v] = src_data[src_idx];
                        }
                    }
@@ -286,3 +286,28 @@ kernel void kernel_cpy_i32_i32(
        dst_data[i00] = src[0];
    }
 }
+
+// Contiguous f32 copy, one work item per float4 over the whole tensor. The kernels above map
+// one workgroup to each row, which leaves a tensor with few long rows on a single compute unit.
+// vload4/vstore4 rather than a float4 cast: these buffers carry an arbitrary 4-byte view offset.
+kernel void kernel_cpy_f32_f32_flat(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        ulong ne,
+        ulong n4
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst  + offsetd);
+
+    const ulong i = get_global_id(0);
+
+    if (i < n4) {
+        vstore4(vload4(i, src0), i, dst);
+    } else if (i == n4) {
+        for (ulong t = n4 * 4; t < ne; ++t) {
+            dst[t] = src0[t];
+        }
+    }
+}
@@ -0,0 +1,871 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_subgroup_uniform_load : enable
+#pragma OPENCL EXTENSION cl_qcom_subgroup_constant_load : enable
+
+#define bool2 uchar2
+#define bool3 uchar3
+#define bool4 uchar4
+
+__constant sampler_t smp_none = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_NONE | CLK_FILTER_NEAREST;
+__constant sampler_t smp_zero = CLK_NORMALIZED_COORDS_FALSE | CLK_ADDRESS_CLAMP | CLK_FILTER_NEAREST;
+
+__kernel void adreno_xmem_attn_q_f32_to_img_scaled(const global void *  src_void,
+                                                   ulong                src_offset,
+                                                   write_only image2d_t dst_image2d,
+                                                   const float          scale,
+                                                   const int            d_head,
+                                                   const int            n_q,
+                                                   const int            n_head,
+                                                   const int            n_head_kv,
+                                                   const int            n_batch,
+                                                   const ulong          src_nb1,
+                                                   const ulong          src_nb2,
+                                                   const ulong          src_nb3) {
+    const int x      = get_global_id(0);
+    const int flat_h = get_global_id(1);
+    const int d      = get_global_id(2);
+
+    const int heads_total = n_head * n_batch;
+    const int kpack       = d_head / 4;
+
+    if (x >= n_q || flat_h >= heads_total || d >= kpack) {
+        return;
+    }
+
+    const int batch = flat_h / n_head;
+    const int head  = flat_h % n_head;
+    const int gqa   = n_head / n_head_kv;
+    const int head_kv = head / gqa;
+    const int head_group = head - head_kv * gqa;
+    const int compact_h = batch * n_head_kv + head_kv;
+    const int compact_x = head_group * n_q + x;
+    const int c     = d * 4;
+
+    const global char *  src_base = (const global char *) src_void + src_offset;
+    const global float * row_ptr  = (const global float *) (src_base + batch * src_nb3 + head * src_nb2 + x * src_nb1);
+
+    half4 out = (half4) (0.0h);
+    out.x     = convert_half(row_ptr[c + 0] * scale);
+    if (c + 1 < d_head) {
+        out.y = convert_half(row_ptr[c + 1] * scale);
+    }
+    if (c + 2 < d_head) {
+        out.z = convert_half(row_ptr[c + 2] * scale);
+    }
+    if (c + 3 < d_head) {
+        out.w = convert_half(row_ptr[c + 3] * scale);
+    }
+
+    write_imageh(dst_image2d, (int2) (compact_x, compact_h * kpack + d), out);
+}
+
+__kernel void adreno_xmem_attn_kv_f32_to_img_gqa(const global void *  src_void,
+                                                 ulong                src_offset,
+                                                 write_only image2d_t dst_image2d,
+                                                 const int            d_head,
+                                                 const int            n_kv,
+                                                 const int            n_kv_padded,
+                                                 const int            n_head_kv,
+                                                 const int            n_batch,
+                                                 const ulong          src_nb1,
+                                                 const ulong          src_nb2,
+                                                 const ulong          src_nb3) {
+    const int x      = get_global_id(0);
+    const int flat_h = get_global_id(1);
+    const int d      = get_global_id(2);
+
+    const int kv_heads_total = n_head_kv * n_batch;
+    const int kpack          = d_head / 4;
+
+    if (x >= n_kv_padded || flat_h >= kv_heads_total || d >= kpack) {
+        return;
+    }
+
+    const int batch   = flat_h / n_head_kv;
+    const int head_kv = flat_h % n_head_kv;
+    const int c       = d * 4;
+
+    half4 out = (half4) (0.0h);
+    if (x < n_kv) {
+        const global char *  src_base = (const global char *) src_void + src_offset;
+        const global float * row_ptr =
+            (const global float *) (src_base + batch * src_nb3 + head_kv * src_nb2 + x * src_nb1);
+        out.x = convert_half(row_ptr[c + 0]);
+        if (c + 1 < d_head) {
+            out.y = convert_half(row_ptr[c + 1]);
+        }
+        if (c + 2 < d_head) {
+            out.z = convert_half(row_ptr[c + 2]);
+        }
+        if (c + 3 < d_head) {
+            out.w = convert_half(row_ptr[c + 3]);
+        }
+    }
+
+    write_imageh(dst_image2d, (int2) (x, flat_h * kpack + d), out);
+}
+
+__kernel void adreno_xmem_attn_kv_f16_to_img_gqa(const global void *  src_void,
+                                                 ulong                src_offset,
+                                                 write_only image2d_t dst_image2d,
+                                                 const int            d_head,
+                                                 const int            n_kv,
+                                                 const int            n_kv_padded,
+                                                 const int            n_head_kv,
+                                                 const int            n_batch,
+                                                 const ulong          src_nb1,
+                                                 const ulong          src_nb2,
+                                                 const ulong          src_nb3) {
+    const int x      = get_global_id(0);
+    const int flat_h = get_global_id(1);
+    const int d      = get_global_id(2);
+
+    const int kv_heads_total = n_head_kv * n_batch;
+    const int kpack          = d_head / 4;
+
+    if (x >= n_kv_padded || flat_h >= kv_heads_total || d >= kpack) {
+        return;
+    }
+
+    const int batch   = flat_h / n_head_kv;
+    const int head_kv = flat_h % n_head_kv;
+    const int c       = d * 4;
+
+    half4 out = (half4) (0.0h);
+    if (x < n_kv) {
+        const global char * src_base = (const global char *) src_void + src_offset;
+        const global half * row_ptr =
+            (const global half *) (src_base + batch * src_nb3 + head_kv * src_nb2 + x * src_nb1);
+        out.x = row_ptr[c + 0];
+        if (c + 1 < d_head) {
+            out.y = row_ptr[c + 1];
+        }
+        if (c + 2 < d_head) {
+            out.z = row_ptr[c + 2];
+        }
+        if (c + 3 < d_head) {
+            out.w = row_ptr[c + 3];
+        }
+    }
+
+    write_imageh(dst_image2d, (int2) (x, flat_h * kpack + d), out);
+}
+
+__kernel void adreno_xmem_attn_img_to_f32(global void *       dst_void,
+                                          ulong               dst_offset,
+                                          read_only image2d_t src_image2d,
+                                          const int           d_head,
+                                          const int           n_q,
+                                          const int           n_head,
+                                          const int           n_head_kv,
+                                          const int           n_batch,
+                                          const ulong         dst_nb1,
+                                          const ulong         dst_nb2,
+                                          const ulong         dst_nb3) {
+    const int x      = get_global_id(0);
+    const int flat_h = get_global_id(1);
+    const int d      = get_global_id(2);
+
+    const int heads_total = n_head * n_batch;
+    const int kpack       = d_head / 4;
+
+    if (x >= n_q || flat_h >= heads_total || d >= kpack) {
+        return;
+    }
+
+    const int batch = flat_h / n_head;
+    const int head  = flat_h % n_head;
+    const int gqa   = n_head / n_head_kv;
+    const int head_kv = head / gqa;
+    const int head_group = head - head_kv * gqa;
+    const int compact_h = batch * n_head_kv + head_kv;
+    const int compact_x = head_group * n_q + x;
+    const int c     = d * 4;
+
+    global char *  dst_base = (global char *) dst_void + dst_offset;
+    global float * row_ptr  = (global float *) (dst_base + batch * dst_nb3 + x * dst_nb2 + head * dst_nb1);
+
+    const half4 in_value = read_imageh(src_image2d, smp_zero, (int2) (compact_x, compact_h * kpack + d));
+    row_ptr[c + 0]       = convert_float(in_value.x);
+    if (c + 1 < d_head) {
+        row_ptr[c + 1] = convert_float(in_value.y);
+    }
+    if (c + 2 < d_head) {
+        row_ptr[c + 2] = convert_float(in_value.z);
+    }
+    if (c + 3 < d_head) {
+        row_ptr[c + 3] = convert_float(in_value.w);
+    }
+}
+
+__kernel void adreno_xmem_attn_k_gather(global half4 *      dst_tensor_buffer,
+                                        read_only image2d_t src_tensor_image2d,
+                                        const int4          shared_int4_0,
+                                        const int4          shared_int4_1) {
+    int X = get_global_id(0);
+    int Y = get_global_id(1);
+    int S = get_global_id(2);
+    if (X >= shared_int4_0.w || Y >= shared_int4_0.y || S >= shared_int4_0.z) {
+        return;
+    }
+    half temps[4];
+    temps[0] = (half) (0.f);
+    temps[1] = (half) (0.f);
+    temps[2] = (half) (0.f);
+    temps[3] = (half) (0.f);
+    for (int i = 0; i < 4; ++i) {
+        int dst_channel = S * 4 + i;
+        if (dst_channel < shared_int4_0.x) {
+            int s_y = Y;
+            int s_x = dst_channel;
+            int s_c = X;
+            {
+                int   slice_coord_TMP  = (s_c) / 4;
+                int   sub_ch_coord_TMP = (s_c) % 4;
+                half4 src_TMP          = read_imageh(src_tensor_image2d, smp_zero,
+                                                     (int2) ((s_x), ((s_y) *shared_int4_1.x + (slice_coord_TMP))));
+                temps[i]               = (half[4]){ src_TMP.x, src_TMP.y, src_TMP.z, src_TMP.w }[sub_ch_coord_TMP];
+            };
+        }
+    }
+    half4 result;
+    result.x                                                                  = temps[0];
+    result.y                                                                  = temps[1];
+    result.z                                                                  = temps[2];
+    result.w                                                                  = temps[3];
+    dst_tensor_buffer[(((S) *shared_int4_0.y + (Y)) * shared_int4_0.w + (X))] = result;
+}
+
+__kernel void adreno_xmem_attn_pack_k(global half4 *             dst_tensor_buffer,
+                                      read_only image1d_buffer_t src_image_buffer,
+                                      const int4                 shared_int4_0,
+                                      const int4                 shared_int4_1,
+                                      const int4                 shared_int4_2) {
+    int linear_index = get_global_id(0);
+    if (linear_index >= shared_int4_0.y) {
+        return;
+    }
+    if (get_global_id(1) != 0) {
+        return;
+    }
+    if (get_global_id(2) != 0) {
+        return;
+    }
+    int   dst_o_sp_i_ogroup = linear_index;
+    int   dst_ogroup        = dst_o_sp_i_ogroup % shared_int4_0.x;
+    int   dst_o_sp_i        = dst_o_sp_i_ogroup / shared_int4_0.x;
+    int   dst_i             = dst_o_sp_i % shared_int4_0.z;
+    int   dst_o_sp          = dst_o_sp_i / shared_int4_0.z;
+    int   dst_sp            = dst_o_sp % shared_int4_1.x;
+    int   dst_o             = dst_o_sp / shared_int4_1.x;
+    int   i_slice           = dst_i;
+    int   o_slice           = dst_o * shared_int4_0.x + dst_ogroup;
+    int   spatial_linear    = dst_sp;
+    int   W                 = spatial_linear % shared_int4_1.y;
+    int   H                 = spatial_linear / shared_int4_1.y;
+    half4 w0                = (half4) (0);
+    half4 w1                = (half4) (0);
+    half4 w2                = (half4) (0);
+    half4 w3                = (half4) (0);
+
+    if (i_slice * 4 < shared_int4_0.w && o_slice < shared_int4_1.w) {
+        w0 = read_imageh(src_image_buffer, (((o_slice) *shared_int4_1.z + (W)) * shared_int4_2.x + (i_slice * 4)));
+    }
+    if (i_slice * 4 + 1 < shared_int4_0.w && o_slice < shared_int4_1.w) {
+        w1 = read_imageh(src_image_buffer, (((o_slice) *shared_int4_1.z + (W)) * shared_int4_2.x + (i_slice * 4 + 1)));
+    }
+    if (i_slice * 4 + 2 < shared_int4_0.w && o_slice < shared_int4_1.w) {
+        w2 = read_imageh(src_image_buffer, (((o_slice) *shared_int4_1.z + (W)) * shared_int4_2.x + (i_slice * 4 + 2)));
+    }
+    if (i_slice * 4 + 3 < shared_int4_0.w && o_slice < shared_int4_1.w) {
+        w3 = read_imageh(src_image_buffer, (((o_slice) *shared_int4_1.z + (W)) * shared_int4_2.x + (i_slice * 4 + 3)));
+    }
+    half4 r0                                = w0;
+    half4 r1                                = w1;
+    half4 r2                                = w2;
+    half4 r3                                = w3;
+    dst_tensor_buffer[linear_index * 4 + 0] = r0;
+    dst_tensor_buffer[linear_index * 4 + 1] = r1;
+    dst_tensor_buffer[linear_index * 4 + 2] = r2;
+    dst_tensor_buffer[linear_index * 4 + 3] = r3;
+}
+
+__attribute__((qcom_max_concurrent_subgroups(12))) __kernel void adreno_xmem_attn_qk_gemm(
+    global half4 *      dst_tensor_buffer,
+    constant half8 *    weights_buffer __attribute__((sub_group_uniform)),
+    constant half8 *    xmem_buffer __attribute__((max_constant_size((6144)))),
+    read_only image2d_t src_tensor_image2d,
+    const int4          shared_int4_0,
+    const int4          shared_int4_1,
+    const int4          shared_int4_2) {
+    int X = get_group_id(1) * get_local_size(0) + get_local_id(0);
+    int Y = get_group_id(2) * get_local_size(1) + get_local_id(1);
+    int Z = get_group_id(0) * get_local_size(2) + get_local_id(2);
+    if (X >= shared_int4_0.z || Y >= shared_int4_0.x) {
+        return;
+    }
+    if (Z * 8 >= shared_int4_0.y) {
+        return;
+    }
+
+    half4 r0      = (half4) (0.f);
+    half4 r1      = (half4) (0.f);
+    half4 r2      = (half4) (0.f);
+    half4 r3      = (half4) (0.f);
+    half4 r4      = (half4) (0.f);
+    half4 r5      = (half4) (0.f);
+    half4 r6      = (half4) (0.f);
+    half4 r7      = (half4) (0.f);
+    int   x_coord = mad24(X, shared_int4_2.y, shared_int4_1.y);
+    int   y_coord = mad24(Y, shared_int4_2.z, shared_int4_1.z);
+    int   coord_x, coord_y, coord_s;
+    int   f_offset = (Z * shared_int4_1.w + Y) * shared_int4_1.x * 32;
+
+    int subgroup_id                   = (int) ((0x1F & qcom_get_physical_sub_group_id()));
+    subgroup_id                       = subgroup_id % 12;
+    int                 c_offset      = mul24(subgroup_id, shared_int4_0.w);
+    __constant half16 * weights_cache = (__constant half16 *) &xmem_buffer[c_offset];
+    coord_y                           = Y;
+    coord_x                           = X;
+    coord_s                           = 0;
+    do {
+        half4 src0 =
+            read_imageh(src_tensor_image2d, smp_zero, (int2) ((coord_x), ((coord_y) *shared_int4_2.x + (coord_s))));
+        coord_s++;
+        half4 src1 =
+            read_imageh(src_tensor_image2d, smp_zero, (int2) ((coord_x), ((coord_y) *shared_int4_2.x + (coord_s))));
+        coord_s++;
+        qcom_sub_group_constant_load8(xmem_buffer, weights_buffer, c_offset, f_offset >> 1, 32);
+        f_offset += 64;
+        qcom_sub_group_sync(QCOM_CLK_CONST_LOAD_SYNC);
+        r0 += src0.x * weights_cache[0].s0123;
+        r0 += src0.y * weights_cache[0].s4567;
+        r0 += src0.z * weights_cache[0].s89ab;
+        r0 += src0.w * weights_cache[0].scdef;
+        r1 += src0.x * weights_cache[1].s0123;
+        r1 += src0.y * weights_cache[1].s4567;
+        r1 += src0.z * weights_cache[1].s89ab;
+        r1 += src0.w * weights_cache[1].scdef;
+        r2 += src0.x * weights_cache[2].s0123;
+        r2 += src0.y * weights_cache[2].s4567;
+        r2 += src0.z * weights_cache[2].s89ab;
+        r2 += src0.w * weights_cache[2].scdef;
+        r3 += src0.x * weights_cache[3].s0123;
+        r3 += src0.y * weights_cache[3].s4567;
+        r3 += src0.z * weights_cache[3].s89ab;
+        r3 += src0.w * weights_cache[3].scdef;
+        r4 += src0.x * weights_cache[4].s0123;
+        r4 += src0.y * weights_cache[4].s4567;
+        r4 += src0.z * weights_cache[4].s89ab;
+        r4 += src0.w * weights_cache[4].scdef;
+        r5 += src0.x * weights_cache[5].s0123;
+        r5 += src0.y * weights_cache[5].s4567;
+        r5 += src0.z * weights_cache[5].s89ab;
+        r5 += src0.w * weights_cache[5].scdef;
+        r6 += src0.x * weights_cache[6].s0123;
+        r6 += src0.y * weights_cache[6].s4567;
+        r6 += src0.z * weights_cache[6].s89ab;
+        r6 += src0.w * weights_cache[6].scdef;
+        r7 += src0.x * weights_cache[7].s0123;
+        r7 += src0.y * weights_cache[7].s4567;
+        r7 += src0.z * weights_cache[7].s89ab;
+        r7 += src0.w * weights_cache[7].scdef;
+        r0 += src1.x * weights_cache[8].s0123;
+        r0 += src1.y * weights_cache[8].s4567;
+        r0 += src1.z * weights_cache[8].s89ab;
+        r0 += src1.w * weights_cache[8].scdef;
+        r1 += src1.x * weights_cache[9].s0123;
+        r1 += src1.y * weights_cache[9].s4567;
+        r1 += src1.z * weights_cache[9].s89ab;
+        r1 += src1.w * weights_cache[9].scdef;
+        r2 += src1.x * weights_cache[10].s0123;
+        r2 += src1.y * weights_cache[10].s4567;
+        r2 += src1.z * weights_cache[10].s89ab;
+        r2 += src1.w * weights_cache[10].scdef;
+        r3 += src1.x * weights_cache[11].s0123;
+        r3 += src1.y * weights_cache[11].s4567;
+        r3 += src1.z * weights_cache[11].s89ab;
+        r3 += src1.w * weights_cache[11].scdef;
+        r4 += src1.x * weights_cache[12].s0123;
+        r4 += src1.y * weights_cache[12].s4567;
+        r4 += src1.z * weights_cache[12].s89ab;
+        r4 += src1.w * weights_cache[12].scdef;
+        r5 += src1.x * weights_cache[13].s0123;
+        r5 += src1.y * weights_cache[13].s4567;
+        r5 += src1.z * weights_cache[13].s89ab;
+        r5 += src1.w * weights_cache[13].scdef;
+        r6 += src1.x * weights_cache[14].s0123;
+        r6 += src1.y * weights_cache[14].s4567;
+        r6 += src1.z * weights_cache[14].s89ab;
+        r6 += src1.w * weights_cache[14].scdef;
+        r7 += src1.x * weights_cache[15].s0123;
+        r7 += src1.y * weights_cache[15].s4567;
+        r7 += src1.z * weights_cache[15].s89ab;
+        r7 += src1.w * weights_cache[15].scdef;
+    } while (coord_s < shared_int4_2.x);
+
+    coord_s = mul24(Z, 8);
+    coord_x = X;
+    coord_y = Y;
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r0);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r1);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r2);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r3);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r4);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r5);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r6);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r7);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image2d, smp_zero, (int2) ((0), ((0) * shared_int4_2.x + (0))));
+        }
+        dst_tensor_buffer[(((coord_s) *shared_int4_0.x + (coord_y)) * shared_int4_0.z + (coord_x))] = res;
+        coord_s++;
+    }
+}
+
+__kernel void adreno_xmem_attn_softmax_reduce_basic(read_only image1d_buffer_t src_tensor_image_buffer,
+                                                    write_only image2d_t       dst_tensor_image2d,
+                                                    const int4                 shared_int4_0,
+                                                    const int4                 shared_int4_1) {
+    int X = get_global_id(0);
+    int Y = get_global_id(1);
+    if (X >= shared_int4_0.z || Y >= shared_int4_0.x) {
+        return;
+    }
+    float sum                     = 0.0f;
+    int   end_channel             = shared_int4_0.w;
+    int   end_slice               = (end_channel + 3) / 4;
+    int   start_channel           = 0;
+    int   start_slice             = start_channel / 4;
+    bool  need_per_channels_check = start_channel % 4 != 0 || end_channel % 4 != 0;
+    float maximum;
+    {
+        int    slice_coord_TMP  = (start_channel) / 4;
+        int    sub_ch_coord_TMP = (start_channel) % 4;
+        float4 src_TMP          = convert_float4(
+            read_imageh(src_tensor_image_buffer, ((slice_coord_TMP) *shared_int4_1.x + (Y)) * shared_int4_1.y + (X)));
+        maximum = (float[4]){ src_TMP.x, src_TMP.y, src_TMP.z, src_TMP.w }[sub_ch_coord_TMP];
+    };
+    for (int d = start_slice; d < end_slice; d += 1) {
+        float4 mask_dot = (float4) (1.f);
+        float4 src =
+            convert_float4(read_imageh(src_tensor_image_buffer, ((d) *shared_int4_1.x + (Y)) * shared_int4_1.y + (X)));
+        if (need_per_channels_check && (d == start_slice || d == end_slice - 1)) {
+            if (d * 4 + 0 < start_channel || d * 4 + 0 >= end_channel) {
+                mask_dot.x = 0.f;
+                src.x      = maximum;
+            }
+            if (d * 4 + 1 < start_channel || d * 4 + 1 >= end_channel) {
+                mask_dot.y = 0.f;
+                src.y      = maximum;
+            }
+            if (d * 4 + 2 < start_channel || d * 4 + 2 >= end_channel) {
+                mask_dot.z = 0.f;
+                src.z      = maximum;
+            }
+            if (d * 4 + 3 < start_channel || d * 4 + 3 >= end_channel) {
+                mask_dot.w = 0.f;
+                src.w      = maximum;
+            }
+        }
+        float new_max = max(src.x, src.y);
+        new_max       = max(new_max, src.z);
+        new_max       = max(new_max, src.w);
+        new_max       = max(new_max, maximum);
+        float scale   = native_exp(maximum - new_max);
+        maximum       = new_max;
+        sum *= scale;
+        float4 exp_res = native_exp(src - maximum);
+        sum += dot(mask_dot, exp_res);
+    }
+    if (!isfinite(maximum) || sum == 0.0f) {
+        write_imageh(dst_tensor_image2d, (int2) (X, Y), (half4) (0.0h));
+        return;
+    }
+    write_imageh(dst_tensor_image2d, (int2) (X, Y),
+                 (half4) (convert_half(1.0f / sum), convert_half(maximum), 0.0h, 0.0h));
+}
+
+__kernel void adreno_xmem_attn_softmax_apply_basic(global half4 *             dst_tensor_buffer,
+                                                   read_only image1d_buffer_t src_tensor_image_buffer,
+                                                   read_only image2d_t        src_tensor_1_image2d,
+                                                   const int4                 shared_int4_0,
+                                                   const int4                 shared_int4_1) {
+    int X = get_global_id(0);
+    int Y = get_global_id(1);
+    int Z = get_global_id(2);
+    if (X >= shared_int4_0.z || Y >= shared_int4_0.x || Z >= shared_int4_0.y) {
+        return;
+    }
+    half4 src = read_imageh(src_tensor_image_buffer, ((Z) *shared_int4_1.x + (Y)) * shared_int4_1.y + (X));
+    {
+        half4 src_final;
+        {
+            {
+                half4 exp_val = read_imageh(src_tensor_1_image2d, smp_zero, (int2) (X, Y));
+                src_final = exp(src - exp_val.y) * exp_val.x;
+                const int k = Z * 4;
+                const int n_kv = shared_int4_1.z;
+                if (k + 0 >= n_kv) {
+                    src_final.x = 0.0h;
+                }
+                if (k + 1 >= n_kv) {
+                    src_final.y = 0.0h;
+                }
+                if (k + 2 >= n_kv) {
+                    src_final.z = 0.0h;
+                }
+                if (k + 3 >= n_kv) {
+                    src_final.w = 0.0h;
+                }
+            }
+        }
+        dst_tensor_buffer[(((Z) *shared_int4_0.x + (Y)) * shared_int4_0.z + (X))] = src_final;
+    };
+}
+
+__kernel void adreno_xmem_attn_mask_scores(global half4 *             dst_score_tensor_buffer,
+                                           read_only image1d_buffer_t src_score_image_buffer,
+                                           const global half *        mask,
+                                           const ulong                mask_offset,
+                                           const int                  q_width,
+                                           const int                  n_q,
+                                           const int                  n_kv,
+                                           const int                  n_kv_padded,
+                                           const int                  kv_heads_total,
+                                           const int                  n_head,
+                                           const int                  n_head_kv,
+                                           const ulong                mask_nb1,
+                                           const ulong                mask_nb2,
+                                           const ulong                mask_nb3,
+                                           const int                  mask_ne2,
+                                           const int                  mask_ne3) {
+    const int X     = get_global_id(0);
+    const int Y     = get_global_id(1);
+    const int Z     = get_global_id(2);
+    const int npack = n_kv_padded / 4;
+    if (X >= q_width || Y >= kv_heads_total || Z >= npack) {
+        return;
+    }
+
+    const int           gqa            = n_head / n_head_kv;
+    const int           head_kv        = Y % n_head_kv;
+    const int           batch          = Y / n_head_kv;
+    const int           head_group     = X / n_q;
+    const int           q              = X - head_group * n_q;
+    const int           head           = head_kv * gqa + head_group;
+    const int           mask_head_idx  = head % mask_ne2;
+    const int           mask_batch_idx = batch % mask_ne3;
+    const global char *  mask_base      = (const global char *) mask + mask_offset;
+    const global half *  mask_row       = (const global half *) (mask_base + mask_batch_idx * mask_nb3 +
+                                                                 mask_head_idx * mask_nb2 + q * mask_nb1);
+
+    const half4 score   = read_imageh(src_score_image_buffer, ((Z * kv_heads_total + Y) * q_width + X));
+    float       vals[4] = {
+        convert_float(score.x),
+        convert_float(score.y),
+        convert_float(score.z),
+        convert_float(score.w),
+    };
+
+    for (int lane = 0; lane < 4; ++lane) {
+        const int k_idx = Z * 4 + lane;
+        if (k_idx >= n_kv) {
+            vals[lane] = -INFINITY;
+        } else {
+            vals[lane] += convert_float(mask_row[k_idx]);
+        }
+    }
+
+    dst_score_tensor_buffer[((Z * kv_heads_total + Y) * q_width + X)] =
+        (half4) (convert_half(vals[0]), convert_half(vals[1]), convert_half(vals[2]), convert_half(vals[3]));
+}
+
+__kernel void adreno_xmem_attn_pack_v(global half4 *      dst_tensor_buffer,
+                                      read_only image2d_t src_image2d,
+                                      const int4          shared_int4_0,
+                                      const int4          shared_int4_1) {
+    int linear_index = get_global_id(0);
+    if (linear_index >= shared_int4_0.y) {
+        return;
+    }
+    if (get_global_id(1) != 0) {
+        return;
+    }
+    if (get_global_id(2) != 0) {
+        return;
+    }
+    int   dst_o_sp_i_ogroup = linear_index;
+    int   dst_ogroup        = dst_o_sp_i_ogroup % shared_int4_0.x;
+    int   dst_o_sp_i        = dst_o_sp_i_ogroup / shared_int4_0.x;
+    int   dst_i             = dst_o_sp_i % shared_int4_0.z;
+    int   dst_o_sp          = dst_o_sp_i / shared_int4_0.z;
+    int   dst_sp            = dst_o_sp % shared_int4_1.x;
+    int   dst_o             = dst_o_sp / shared_int4_1.x;
+    int   i_slice           = dst_i;
+    int   o_slice           = dst_o * shared_int4_0.x + dst_ogroup;
+    int   spatial_linear    = dst_sp;
+    int   W                 = spatial_linear % shared_int4_1.y;
+    int   H                 = spatial_linear / shared_int4_1.y;
+    half4 w0                = (half4) (0);
+    half4 w1                = (half4) (0);
+    half4 w2                = (half4) (0);
+    half4 w3                = (half4) (0);
+
+    if (i_slice * 4 < shared_int4_0.w && o_slice < shared_int4_1.z) {
+        w0 = read_imageh(src_image2d, smp_zero, (int2) ((i_slice * 4), ((W) *shared_int4_1.z + (o_slice))));
+    }
+    if (i_slice * 4 + 1 < shared_int4_0.w && o_slice < shared_int4_1.z) {
+        w1 = read_imageh(src_image2d, smp_zero, (int2) ((i_slice * 4 + 1), ((W) *shared_int4_1.z + (o_slice))));
+    }
+    if (i_slice * 4 + 2 < shared_int4_0.w && o_slice < shared_int4_1.z) {
+        w2 = read_imageh(src_image2d, smp_zero, (int2) ((i_slice * 4 + 2), ((W) *shared_int4_1.z + (o_slice))));
+    }
+    if (i_slice * 4 + 3 < shared_int4_0.w && o_slice < shared_int4_1.z) {
+        w3 = read_imageh(src_image2d, smp_zero, (int2) ((i_slice * 4 + 3), ((W) *shared_int4_1.z + (o_slice))));
+    }
+    half4 r0                                = w0;
+    half4 r1                                = w1;
+    half4 r2                                = w2;
+    half4 r3                                = w3;
+    dst_tensor_buffer[linear_index * 4 + 0] = r0;
+    dst_tensor_buffer[linear_index * 4 + 1] = r1;
+    dst_tensor_buffer[linear_index * 4 + 2] = r2;
+    dst_tensor_buffer[linear_index * 4 + 3] = r3;
+}
+
+__attribute__((qcom_max_concurrent_subgroups(12))) __kernel void adreno_xmem_attn_pv_gemm(
+    constant half8 *           weights_buffer __attribute__((sub_group_uniform)),
+    constant half8 *           xmem_buffer __attribute__((max_constant_size((6144)))),
+    read_only image1d_buffer_t src_tensor_image_buffer,
+    write_only image2d_t       dst_tensor_image2d,
+    const int4                 shared_int4_0,
+    const int4                 shared_int4_1,
+    const int4                 shared_int4_2,
+    const int4                 shared_int4_3) {
+    int X = get_group_id(1) * get_local_size(0) + get_local_id(0);
+    int Y = get_group_id(2) * get_local_size(1) + get_local_id(1);
+    int Z = get_group_id(0) * get_local_size(2) + get_local_id(2);
+    if (X >= shared_int4_0.z || Y >= shared_int4_0.x) {
+        return;
+    }
+    if (Z * 8 >= shared_int4_0.y) {
+        return;
+    }
+
+    half4 r0      = (half4) (0.f);
+    half4 r1      = (half4) (0.f);
+    half4 r2      = (half4) (0.f);
+    half4 r3      = (half4) (0.f);
+    half4 r4      = (half4) (0.f);
+    half4 r5      = (half4) (0.f);
+    half4 r6      = (half4) (0.f);
+    half4 r7      = (half4) (0.f);
+    int   x_coord = mad24(X, shared_int4_2.w, shared_int4_1.y);
+    int   y_coord = mad24(Y, shared_int4_3.x, shared_int4_1.z);
+    int   coord_x, coord_y, coord_s;
+    int   f_offset = (Z * shared_int4_1.w + Y) * shared_int4_1.x * 32;
+
+    int subgroup_id                   = (int) ((0x1F & qcom_get_physical_sub_group_id()));
+    subgroup_id                       = subgroup_id % 12;
+    int                 c_offset      = mul24(subgroup_id, shared_int4_0.w);
+    __constant half16 * weights_cache = (__constant half16 *) &xmem_buffer[c_offset];
+    coord_y                           = Y;
+    coord_x                           = X;
+    int addr                          = (((0) * shared_int4_1.w + (coord_y)) * shared_int4_2.z + (coord_x));
+    int dz                            = shared_int4_2.x;
+    coord_s                           = 0;
+    do {
+        half4 src0 = read_imageh(src_tensor_image_buffer, addr);
+        addr += dz;
+        coord_s++;
+        half4 src1 = read_imageh(src_tensor_image_buffer, addr);
+        addr += dz;
+        coord_s++;
+        qcom_sub_group_constant_load8(xmem_buffer, weights_buffer, c_offset, f_offset >> 1, 32);
+        f_offset += 64;
+        qcom_sub_group_sync(QCOM_CLK_CONST_LOAD_SYNC);
+        r0 += src0.x * weights_cache[0].s0123;
+        r0 += src0.y * weights_cache[0].s4567;
+        r0 += src0.z * weights_cache[0].s89ab;
+        r0 += src0.w * weights_cache[0].scdef;
+        r1 += src0.x * weights_cache[1].s0123;
+        r1 += src0.y * weights_cache[1].s4567;
+        r1 += src0.z * weights_cache[1].s89ab;
+        r1 += src0.w * weights_cache[1].scdef;
+        r2 += src0.x * weights_cache[2].s0123;
+        r2 += src0.y * weights_cache[2].s4567;
+        r2 += src0.z * weights_cache[2].s89ab;
+        r2 += src0.w * weights_cache[2].scdef;
+        r3 += src0.x * weights_cache[3].s0123;
+        r3 += src0.y * weights_cache[3].s4567;
+        r3 += src0.z * weights_cache[3].s89ab;
+        r3 += src0.w * weights_cache[3].scdef;
+        r4 += src0.x * weights_cache[4].s0123;
+        r4 += src0.y * weights_cache[4].s4567;
+        r4 += src0.z * weights_cache[4].s89ab;
+        r4 += src0.w * weights_cache[4].scdef;
+        r5 += src0.x * weights_cache[5].s0123;
+        r5 += src0.y * weights_cache[5].s4567;
+        r5 += src0.z * weights_cache[5].s89ab;
+        r5 += src0.w * weights_cache[5].scdef;
+        r6 += src0.x * weights_cache[6].s0123;
+        r6 += src0.y * weights_cache[6].s4567;
+        r6 += src0.z * weights_cache[6].s89ab;
+        r6 += src0.w * weights_cache[6].scdef;
+        r7 += src0.x * weights_cache[7].s0123;
+        r7 += src0.y * weights_cache[7].s4567;
+        r7 += src0.z * weights_cache[7].s89ab;
+        r7 += src0.w * weights_cache[7].scdef;
+        r0 += src1.x * weights_cache[8].s0123;
+        r0 += src1.y * weights_cache[8].s4567;
+        r0 += src1.z * weights_cache[8].s89ab;
+        r0 += src1.w * weights_cache[8].scdef;
+        r1 += src1.x * weights_cache[9].s0123;
+        r1 += src1.y * weights_cache[9].s4567;
+        r1 += src1.z * weights_cache[9].s89ab;
+        r1 += src1.w * weights_cache[9].scdef;
+        r2 += src1.x * weights_cache[10].s0123;
+        r2 += src1.y * weights_cache[10].s4567;
+        r2 += src1.z * weights_cache[10].s89ab;
+        r2 += src1.w * weights_cache[10].scdef;
+        r3 += src1.x * weights_cache[11].s0123;
+        r3 += src1.y * weights_cache[11].s4567;
+        r3 += src1.z * weights_cache[11].s89ab;
+        r3 += src1.w * weights_cache[11].scdef;
+        r4 += src1.x * weights_cache[12].s0123;
+        r4 += src1.y * weights_cache[12].s4567;
+        r4 += src1.z * weights_cache[12].s89ab;
+        r4 += src1.w * weights_cache[12].scdef;
+        r5 += src1.x * weights_cache[13].s0123;
+        r5 += src1.y * weights_cache[13].s4567;
+        r5 += src1.z * weights_cache[13].s89ab;
+        r5 += src1.w * weights_cache[13].scdef;
+        r6 += src1.x * weights_cache[14].s0123;
+        r6 += src1.y * weights_cache[14].s4567;
+        r6 += src1.z * weights_cache[14].s89ab;
+        r6 += src1.w * weights_cache[14].scdef;
+        r7 += src1.x * weights_cache[15].s0123;
+        r7 += src1.y * weights_cache[15].s4567;
+        r7 += src1.z * weights_cache[15].s89ab;
+        r7 += src1.w * weights_cache[15].scdef;
+    } while (coord_s < shared_int4_2.y);
+
+    coord_s = mul24(Z, 8);
+    coord_x = X;
+    coord_y = Y;
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r0);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r1);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r2);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r3);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r4);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r5);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r6);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+    if (coord_s < shared_int4_0.y) {
+        half4 res = convert_half4(r7);
+        if (coord_s < 0) {
+            res += read_imageh(src_tensor_image_buffer, ((0) * shared_int4_1.w + (0)) * shared_int4_2.z + (0));
+        }
+        write_imageh(dst_tensor_image2d, (int2) ((coord_x), ((coord_y) *shared_int4_0.y + (coord_s))), res);
+        coord_s++;
+    }
+}
@@ -0,0 +1,85 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// Extended elementwise unary ops, same variant shape as abs.cl:
+//   f32, f32_4 (vec4), f16, f16_4 (vec4), f32_nc, f16_nc (stride-addressed).
+//
+//   sgn, step, elu, hardswish, hardsigmoid, floor, ceil, round, trunc.
+//
+// Semantics match the ggml CPU reference (ggml.c). Values are computed in float
+// (the f16 variants read/write half and convert), so the conditional ops match
+// the CPU bit-for-bit within tolerance. SEXPR is the scalar form, VEXPR the
+// float4 form (vector ternaries need select()).
+//------------------------------------------------------------------------------
+
+#define UNARY_EXT(NAME, SEXPR, VEXPR)                                           \
+kernel void kernel_##NAME##_f32(                                               \
+        global const float * src0, ulong offset0,                             \
+        global       float * dst,  ulong offsetd) {                           \
+    src0 = (global float*)((global char*)src0 + offset0);                     \
+    dst  = (global float*)((global char*)dst  + offsetd);                     \
+    float x = src0[get_global_id(0)];                                         \
+    dst[get_global_id(0)] = (SEXPR);                                          \
+}                                                                             \
+kernel void kernel_##NAME##_f32_4(                                            \
+        global const float4 * src0, ulong offset0,                           \
+        global       float4 * dst,  ulong offsetd) {                         \
+    src0 = (global float4*)((global char*)src0 + offset0);                   \
+    dst  = (global float4*)((global char*)dst  + offsetd);                   \
+    float4 x = src0[get_global_id(0)];                                       \
+    dst[get_global_id(0)] = (VEXPR);                                         \
+}                                                                             \
+kernel void kernel_##NAME##_f16(                                             \
+        global const half * src0, ulong offset0,                            \
+        global       half * dst,  ulong offsetd) {                          \
+    src0 = (global half*)((global char*)src0 + offset0);                    \
+    dst  = (global half*)((global char*)dst  + offsetd);                    \
+    float x = src0[get_global_id(0)];                                       \
+    dst[get_global_id(0)] = (SEXPR);                                        \
+}                                                                            \
+kernel void kernel_##NAME##_f16_4(                                          \
+        global const half4 * src0, ulong offset0,                          \
+        global       half4 * dst,  ulong offsetd) {                        \
+    src0 = (global half4*)((global char*)src0 + offset0);                  \
+    dst  = (global half4*)((global char*)dst  + offsetd);                  \
+    float4 x = convert_float4(src0[get_global_id(0)]);                     \
+    dst[get_global_id(0)] = convert_half4(VEXPR);                          \
+}                                                                           \
+kernel void kernel_##NAME##_f32_nc(                                         \
+        global const char * src0, ulong offset0,                          \
+        global       char * dst,  ulong offsetd,                          \
+        int ne00, ulong nb00, ulong nb01, ulong nb02, ulong nb03,         \
+        ulong nb0, ulong nb1, ulong nb2, ulong nb3) {                     \
+    src0 = src0 + offset0; dst = dst + offsetd;                            \
+    const int i3 = get_group_id(2);                                       \
+    const int i2 = get_group_id(1);                                       \
+    const int i1 = get_group_id(0);                                       \
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {  \
+        float x = *(global const float *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); \
+        *(global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0) = (SEXPR);            \
+    }                                                                     \
+}                                                                         \
+kernel void kernel_##NAME##_f16_nc(                                       \
+        global const char * src0, ulong offset0,                        \
+        global       char * dst,  ulong offsetd,                        \
+        int ne00, ulong nb00, ulong nb01, ulong nb02, ulong nb03,       \
+        ulong nb0, ulong nb1, ulong nb2, ulong nb3) {                   \
+    src0 = src0 + offset0; dst = dst + offsetd;                          \
+    const int i3 = get_group_id(2);                                     \
+    const int i2 = get_group_id(1);                                     \
+    const int i1 = get_group_id(0);                                     \
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {\
+        float x = *(global const half *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); \
+        *(global half *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0) = (SEXPR);            \
+    }                                                                   \
+}
+
+UNARY_EXT(sgn,          sign(x),                                           sign(x))
+UNARY_EXT(step,         x > 0.0f ? 1.0f : 0.0f,                            select((float4)0.0f, (float4)1.0f, x > 0.0f))
+UNARY_EXT(elu,          x > 0.0f ? x : expm1(x),                           select(expm1(x), x, x > 0.0f))
+UNARY_EXT(hardswish,    x * fmin(1.0f, fmax(0.0f, (x + 3.0f) / 6.0f)),     x * fmin((float4)1.0f, fmax((float4)0.0f, (x + 3.0f) / 6.0f)))
+UNARY_EXT(hardsigmoid,  fmin(1.0f, fmax(0.0f, (x + 3.0f) / 6.0f)),         fmin((float4)1.0f, fmax((float4)0.0f, (x + 3.0f) / 6.0f)))
+UNARY_EXT(floor,        floor(x),                                          floor(x))
+UNARY_EXT(ceil,         ceil(x),                                           ceil(x))
+UNARY_EXT(round,        round(x),                                          round(x))
+UNARY_EXT(trunc,        trunc(x),                                          trunc(x))
@@ -1091,6 +1091,10 @@ static ggml_openvino_op_support is_op_supported_case(const ggml_tensor * op) {
        if (op->ne[3] != 1) {
            return {false, "GET_ROWS/SET_ROWS with ne[3] != 1 (ne[3]=" + std::to_string(op->ne[3]) + ") is not supported"};
        }
+        if (op->op == GGML_OP_GET_ROWS && ggml_is_quantized(op->src[0]->type) &&
+            op->src[0]->view_src != nullptr && op->src[0]->view_offs != 0) {
+            return {false, "GET_ROWS with a nonzero quantized src0 view offset is not supported"};
+        }
        if (op->op == GGML_OP_GET_ROWS && ggml_openvino_get_device_name() == "GPU" &&
            op->src[0]->type == GGML_TYPE_BF16) {
            return {false, "GET_ROWS with BF16 src0 is not supported on GPU"};
@@ -94,7 +94,7 @@ static bool ggml_sycl_use_level_zero_device_alloc(sycl::queue &q) {

 // Use Level Zero zeMemAllocDevice to avoid sycl::malloc_device triggering
 // DMA-buf/TTM system RAM staging in the xe kernel driver during multi-GPU inference.
-void * ggml_sycl_malloc_device(size_t size, sycl::queue &q) {
+void * ggml_sycl_malloc_device(size_t size, sycl::queue &q, ggml_sycl_mem_type type) {
 #ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
    if (ggml_sycl_use_level_zero_device_alloc(q)) {
        void *ptr = nullptr;
@@ -117,16 +117,25 @@ void * ggml_sycl_malloc_device(size_t size, sycl::queue &q) {
 #endif
        ze_result_t r = zeMemAllocDevice(ze_ctx, &alloc_desc, size, 64, ze_dev, &ptr);
        if (r == ZE_RESULT_SUCCESS && ptr) {
+            ggml_sycl_memtrace_add(type, ptr, size);
            return ptr;
        }
+        ggml_sycl_memtrace_fail(type, size);
        return nullptr;
    }
 #endif
-    return sycl::malloc_device(size, q);
+    void * ptr = sycl::malloc_device(size, q);
+    if (ptr == nullptr) {
+        ggml_sycl_memtrace_fail(type, size);
+        return nullptr;
+    }
+    ggml_sycl_memtrace_add(type, ptr, size);
+    return ptr;
 }

 void ggml_sycl_free_device(void *ptr, sycl::queue &q) {
    if (!ptr) return;
+    ggml_sycl_memtrace_del(ptr);
 #ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
    if (ggml_sycl_use_level_zero_device_alloc(q)) {
        auto ze_ctx = sycl::get_native<sycl::backend::ext_oneapi_level_zero>(q.get_context());
@@ -27,6 +27,7 @@
 #include "type.hpp"
 #include "sycl_hw.hpp"
 #include "fattn-buffers.hpp"
+#include "memtrace.hpp"

 namespace syclexp = sycl::ext::oneapi::experimental;

@@ -69,6 +70,8 @@ extern int g_ggml_sycl_dev2dev_memcpy;
 extern int g_ggml_sycl_fa_onednn;
 extern int g_ggml_sycl_fa_onednn_max_kv;
 extern int g_ggml_sycl_enable_mkl_fa;
+extern int g_ggml_sycl_memtrace;
+extern int g_ggml_sycl_memtrace_step;


 #define CHECK_TRY_ERROR(expr)                                            \
@@ -318,7 +321,8 @@ struct ggml_tensor_extra_gpu {
 };

 extern int g_ggml_sycl_use_level_zero_api;
-void * ggml_sycl_malloc_device(size_t size, sycl::queue &q);
+void * ggml_sycl_malloc_device(size_t size, sycl::queue &q,
+                               ggml_sycl_mem_type type = GGML_SYCL_MEM_DIRECT);
 void ggml_sycl_free_device(void *ptr, sycl::queue &q);

 void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> streams={});
@@ -21,6 +21,7 @@ sycl::half * ggml_sycl_fattn_kv_buffers::kv_buffer::ensure_half(size_t n_elems)

    if (ptr) {
        SYCL_CHECK(CHECK_TRY_ERROR(qptr->wait()));
+        ggml_sycl_memtrace_del(ptr);
        SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr)));
        ptr = nullptr;
        capacity = 0;
@@ -38,11 +39,13 @@ sycl::half * ggml_sycl_fattn_kv_buffers::kv_buffer::ensure_half(size_t n_elems)

    if (!dev_ptr) {
        GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, cap);
+        ggml_sycl_memtrace_fail(GGML_SYCL_MEM_FATTN_KV, cap);
        GGML_ABORT("fattn buffer alloc failed");
    }

    ptr = static_cast<sycl::half *>(dev_ptr);
    capacity = cap;
+    ggml_sycl_memtrace_add(GGML_SYCL_MEM_FATTN_KV, ptr, cap);
    return ptr;
 }

@@ -51,6 +54,7 @@ ggml_sycl_fattn_kv_buffers::kv_buffer::~kv_buffer() {
    GGML_LOG_INFO("ggml_sycl_fattn_kv_buffer[%d]: %.2f MiB\n", device, capacity / 1024.0 / 1024.0);
 #endif
    if (ptr) {
+        ggml_sycl_memtrace_del(ptr);
        SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr)));
    }
 }
@@ -1,6 +1,50 @@
 #include "fwht.hpp"

 #include <cmath>
+#define P 1.0f
+#define N -1.0f
+
+// constant Hadamard matrix via Paley I construction
+static constexpr float H12[12][12] = {
+    { P, P, P, P, P, P, P, P, P, P, P, P },
+    { P, N, P, N, P, P, P, N, N, N, P, N },
+    { P, N, N, P, N, P, P, P, N, N, N, P },
+    { P, P, N, N, P, N, P, P, P, N, N, N },
+    { P, N, P, N, N, P, N, P, P, P, N, N },
+    { P, N, N, P, N, N, P, N, P, P, P, N },
+    { P, N, N, N, P, N, N, P, N, P, P, P },
+    { P, P, N, N, N, P, N, N, P, N, P, P },
+    { P, P, P, N, N, N, P, N, N, P, N, P },
+    { P, P, P, P, N, N, N, P, N, N, P, N },
+    { P, N, P, P, P, N, N, N, P, N, N, P },
+    { P, P, N, P, P, P, N, N, N, P, N, N }
+};
+
+static constexpr float H20[20][20] = {
+    { P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P, P },
+    { P, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N },
+    { P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P },
+    { P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P },
+    { P, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N },
+    { P, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N, N },
+    { P, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N, N },
+    { P, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P, N },
+    { P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N, P },
+    { P, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P, N },
+    { P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N, P },
+    { P, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P, N },
+    { P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P, P },
+    { P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P, P },
+    { P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P, P },
+    { P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N, P },
+    { P, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N, N },
+    { P, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P, N },
+    { P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N, P },
+    { P, P, N, N, P, P, P, P, N, P, N, P, N, N, N, N, P, P, N, N }
+};
+
+#undef P
+#undef N

 template <int N>
 static void fwht_kernel(const float * __restrict__ src, float * __restrict__ dst, const int64_t n_rows,
@@ -80,6 +124,122 @@ static void launch_fwht(const float * src, float * dst, const int64_t n_rows, co
                         });
 }

+template <int N, int m>
+static void kronecker_kernel(const float * __restrict__ src,
+                             float * __restrict__ dst,
+                             const int64_t            n_rows,
+                             const float              scale,
+                             const sycl::nd_item<2> & item) {
+    static_assert(m == 12 || m == 20, "block size has to be 12 or 20.");
+
+    const sycl::sub_group sg = item.get_sub_group();
+
+    const int64_t r = item.get_global_id(0);
+    if (r >= n_rows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    constexpr int blocks_per_group = N / m;
+    constexpr int el_w             = blocks_per_group / WARP_SIZE;
+    static_assert(el_w >= 1 && blocks_per_group % WARP_SIZE == 0, "blocks_per_group must be a multiple of WARP_SIZE");
+    float     reg[el_w * m];
+    const int lane = sg.get_local_linear_id();
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        const int b_idx = i * WARP_SIZE + lane;
+
+#pragma unroll
+        for (int j = 0; j < m; ++j) {
+            reg[i * m + j] = src[b_idx * m + j] * scale;
+        }
+    }
+
+#pragma unroll
+    for (int b = 0; b < el_w; ++b) {
+        float z[m] = { 0.0f };
+
+#pragma unroll
+        for (int i = 0; i < m; ++i) {
+#pragma unroll
+            for (int j = 0; j < m; ++j) {
+                const float h = (m == 12 ? H12[j][i] : H20[j][i]);
+                z[i] += reg[b * m + j] * h;
+            }
+        }
+
+#pragma unroll
+        for (int i = 0; i < m; ++i) {
+            reg[b * m + i] = z[i];
+        }
+    }
+
+#pragma unroll
+    for (int h = 1; h < WARP_SIZE; h *= 2) {
+#pragma unroll
+        for (int j = 0; j < el_w; ++j) {
+#pragma unroll
+            for (int k = 0; k < m; ++k) {
+                const float val  = reg[j * m + k];
+                const float val2 = dpct::permute_sub_group_by_xor(sg, val, h, WARP_SIZE);
+
+                reg[j * m + k] = (lane & h) == 0 ? val + val2 : val2 - val;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int h = WARP_SIZE; h < blocks_per_group; h *= 2) {
+        const int step = h / WARP_SIZE;
+#pragma unroll
+        for (int j = 0; j < el_w; j += 2 * step) {
+#pragma unroll
+            for (int s = 0; s < step; ++s) {
+#pragma unroll
+                for (int k = 0; k < m; ++k) {
+                    const float x = reg[(j + s) * m + k];
+                    const float y = reg[(j + s + step) * m + k];
+
+                    reg[(j + s) * m + k]        = x + y;
+                    reg[(j + s + step) * m + k] = x - y;
+                }
+            }
+        }
+    }
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        const int b_idx = i * WARP_SIZE + lane;
+#pragma unroll
+        for (int k = 0; k < m; ++k) {
+            dst[b_idx * m + k] = reg[i * m + k];
+        }
+    }
+}
+
+template <int N, int m>
+static void launch_kronecker(const float *   src,
+                             float *         dst,
+                             const int64_t   n_rows,
+                             const float     scale,
+                             dpct::queue_ptr stream) {
+    constexpr int rows_per_block = 4;
+
+    const int64_t num_blocks = (n_rows + rows_per_block - 1) / rows_per_block;
+
+    // dim 1 is the fastest-varying, so a sub-group is exactly one row's WARP_SIZE lanes.
+    const sycl::range<2> global(num_blocks * rows_per_block, WARP_SIZE);
+    const sycl::range<2> local(rows_per_block, WARP_SIZE);
+
+    stream->parallel_for(sycl::nd_range<2>(global, local),
+                         [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             kronecker_kernel<N, m>(src, dst, n_rows, scale, item);
+                         });
+}
+
 bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst) {
    if (src->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
        return false;
@@ -113,6 +273,18 @@ bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src,
        case 512:
            launch_fwht<512>(src_d, dst_d, rows, scale, stream);
            return true;
+        case 384:
+            launch_kronecker<384, 12>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 768:
+            launch_kronecker<768, 12>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 640:
+            launch_kronecker<640, 20>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 1280:
+            launch_kronecker<1280, 20>(src_d, dst_d, rows, scale, stream);
+            return true;
        default:
            return false;
    }
@@ -97,6 +97,8 @@ int g_ggml_sycl_enable_dnn = 1;
 int g_ggml_sycl_fa_onednn = 1;
 int g_ggml_sycl_fa_onednn_max_kv = 0;
 int g_ggml_sycl_enable_mkl_fa = 1;
+int g_ggml_sycl_memtrace = 0;
+int g_ggml_sycl_memtrace_step = 64;
 int g_ggml_sycl_enable_vmm = 1;
 int g_ggml_sycl_enable_fusion = 1;
 int g_ggml_sycl_enable_esimd = 1;
@@ -335,6 +337,8 @@ static void ggml_check_sycl() try {
        g_ggml_sycl_fa_onednn = ggml_sycl_get_env("GGML_SYCL_FA_ONEDNN", 1);
        g_ggml_sycl_fa_onednn_max_kv = ggml_sycl_get_env("GGML_SYCL_FA_ONEDNN_MAX_KV", 0);
        g_ggml_sycl_enable_mkl_fa = ggml_sycl_get_env("GGML_SYCL_ENABLE_MKL_FA", 1);
+        g_ggml_sycl_memtrace = ggml_sycl_get_env("GGML_SYCL_MEMTRACE", 0);
+        g_ggml_sycl_memtrace_step = ggml_sycl_get_env("GGML_SYCL_MEMTRACE_STEP", 64);
        g_ggml_sycl_enable_vmm = ggml_sycl_get_env("GGML_SYCL_ENABLE_VMM", 1);
        g_ggml_sycl_enable_fusion = ggml_sycl_get_env("GGML_SYCL_ENABLE_FUSION", 1);
        g_ggml_sycl_enable_esimd = ggml_sycl_get_env("GGML_SYCL_ENABLE_ESIMD", 1);
@@ -421,6 +425,8 @@ static void ggml_check_sycl() try {
 #endif
        GGML_LOG_INFO("  GGML_SYCL_FA_ONEDNN_MAX_KV: %d\n", g_ggml_sycl_fa_onednn_max_kv);
        GGML_LOG_INFO("  GGML_SYCL_ENABLE_MKL_FA: %d\n", g_ggml_sycl_enable_mkl_fa);
+        GGML_LOG_INFO("  GGML_SYCL_MEMTRACE: %d\n", g_ggml_sycl_memtrace);
+        GGML_LOG_INFO("  GGML_SYCL_MEMTRACE_STEP: %d\n", g_ggml_sycl_memtrace_step);
 #ifdef SYCL_FLASH_ATTN
        GGML_LOG_INFO("  GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention);
 #else
@@ -964,7 +970,7 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
            return nullptr;
        }
    } else {
-        SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)ggml_sycl_malloc_device(size, *stream)));
+        SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)ggml_sycl_malloc_device(size, *stream, GGML_SYCL_MEM_BUFFER)));
        if (!dev_ptr) {
          GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
          return nullptr;
@@ -1217,7 +1223,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
        ggml_sycl_set_device(i);
        const queue_ptr stream = ctx->streams[i];
        char * buf;
-        SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)ggml_sycl_malloc_device(size, *stream)));
+        SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)ggml_sycl_malloc_device(size, *stream, GGML_SYCL_MEM_BUFFER)));
        if (!buf) {
            char err_buf[1024];
            snprintf(err_buf, 1023, "%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
@@ -1697,7 +1703,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
        void * ptr;
        size_t look_ahead_size = (size_t) (1.05 * size);

-        SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *)ggml_sycl_malloc_device(look_ahead_size, *qptr)));
+        SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *)ggml_sycl_malloc_device(look_ahead_size, *qptr, GGML_SYCL_MEM_POOL_LEG)));
        if (!ptr) {
            GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
            return nullptr;
@@ -1786,6 +1792,13 @@ struct ggml_sycl_pool_vmm : public ggml_sycl_pool {

            GGML_ASSERT(pool_size + reserve_size <= SYCL_POOL_VMM_MAX_SIZE);

+            if (ggml_sycl_memtrace_enabled()) {
+                GGML_LOG_INFO(GGML_SYCL_MEMTRACE_TAG " pool_vmm[%d] committing %5zu MiB (pool %5zu -> %5zu MiB)\n",
+                              device, reserve_size / (1024 * 1024), pool_size / (1024 * 1024),
+                              (pool_size + reserve_size) / (1024 * 1024));
+                ggml_sycl_memtrace_report("before pool_vmm commit");
+            }
+
            // allocate more physical memory
            std::optional<sycl::ext::oneapi::experimental::physical_mem> phys;
            SYCL_CHECK(CHECK_TRY_ERROR(phys.emplace(dev, ctx, reserve_size)));
@@ -1811,6 +1824,7 @@ struct ggml_sycl_pool_vmm : public ggml_sycl_pool {

            // add to the pool
            pool_size += reserve_size;
+            ggml_sycl_memtrace_add(GGML_SYCL_MEM_POOL_VMM, map_ptr, reserve_size);

 #ifdef DEBUG_SYCL_MALLOC
            GGML_LOG_INFO("sycl pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
@@ -4039,7 +4053,9 @@ static inline void * sycl_ext_malloc_device(dpct::queue_ptr stream, size_t size)
    bool use_async = g_ggml_sycl_use_async_mem_op;
 #if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
    if (use_async) {
-        return syclex::async_malloc(*stream, sycl::usm::alloc::device, size);
+        void * ptr = syclex::async_malloc(*stream, sycl::usm::alloc::device, size);
+        ggml_sycl_memtrace_add(GGML_SYCL_MEM_ASYNC, ptr, size);
+        return ptr;
    }
 #else
    // If async allocation extension is not available, use_async should always be false.
@@ -4052,6 +4068,7 @@ static inline void sycl_ext_free(dpct::queue_ptr stream, void * ptr) {
    bool use_async = g_ggml_sycl_use_async_mem_op;
 #if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
    if (use_async) {
+        ggml_sycl_memtrace_del(ptr);
        syclex::async_free(*stream, ptr);
        return;
    }
@@ -4841,6 +4858,78 @@ static bool ggml_sycl_mul_mat_glu_mmvq_fused(ggml_backend_sycl_context & ctx, gg
                                               /*stride_col_dst=*/(int) glu->ne[0], stream);
 }

+// Batch the run of consecutive L2_NORM siblings starting at node_idx into one launch.
+// Returns the number of extra graph nodes consumed, or 0 if the run is shorter than two
+// (the caller then runs the norm through the per-tensor kernel).
+static int ggml_sycl_l2_norm_batch_fused(ggml_backend_sycl_context & ctx, ggml_cgraph * cgraph, int node_idx) {
+    const ggml_tensor * node = cgraph->nodes[node_idx];
+    if (ggml_sycl_info().device_count != 1 || node->type != GGML_TYPE_F32 ||
+        node->src[0]->type != GGML_TYPE_F32 || node->src[0]->ne[0] >= 1024) {
+        return 0;
+    }
+
+    ggml_tensor * batch[GGML_SYCL_L2_BATCH_MAX];
+    int           count = 0;
+    int           last  = node_idx;
+    float         eps0;
+    memcpy(&eps0, node->op_params, sizeof(float));
+
+    // Conservative aliasing test: the batched norms run concurrently in one kernel,
+    // so none may read what another writes, and none may write where another writes.
+    auto overlaps = [](const ggml_tensor * a, const ggml_tensor * b) {
+        const char * ab = (const char *) a->data;
+        const char * bb = (const char *) b->data;
+        return ab < bb + ggml_nbytes(b) && bb < ab + ggml_nbytes(a);
+    };
+
+    for (int j = node_idx; j < cgraph->n_nodes && count < GGML_SYCL_L2_BATCH_MAX; ++j) {
+        ggml_tensor * nj = cgraph->nodes[j];
+        if (ggml_is_empty(nj) || nj->op == GGML_OP_RESHAPE || nj->op == GGML_OP_TRANSPOSE ||
+            nj->op == GGML_OP_VIEW || nj->op == GGML_OP_PERMUTE || nj->op == GGML_OP_NONE ||
+            (nj->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;  // not a launch; cannot break a run of adjacent norms
+        }
+        if (nj->op != GGML_OP_L2_NORM || nj->type != GGML_TYPE_F32 ||
+            nj->src[0]->type != GGML_TYPE_F32 || !ggml_are_same_shape(nj, node) ||
+            !ggml_are_same_shape(nj->src[0], node->src[0])) {
+            break;  // any other launch ends the run
+        }
+        bool same_nb = true;
+        for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+            if (nj->nb[d] != node->nb[d] || nj->src[0]->nb[d] != node->src[0]->nb[d]) {
+                same_nb = false;
+                break;
+            }
+        }
+        if (!same_nb) {
+            break;  // one nb[] stride set is shared by the whole batch
+        }
+        float epsj;
+        memcpy(&epsj, nj->op_params, sizeof(float));
+        if (epsj != eps0) {
+            break;  // eps mismatch ends the run
+        }
+        bool indep = true;
+        for (int k = 0; k < count; ++k) {
+            if (overlaps(nj->src[0], batch[k]) || overlaps(nj, batch[k])) {
+                indep = false;
+                break;
+            }
+        }
+        if (!indep) {
+            break;  // an overlapping tensor would race inside one launch
+        }
+        batch[count++] = nj;
+        last           = j;
+    }
+    if (count < 2) {
+        return 0;  // a lone norm falls through to the per-tensor kernel
+    }
+    ggml_sycl_l2_norm_batch(ctx, batch, count);
+    return last - node_idx;
+}
+
+
 __dpct_inline__ static void k_copy_src1_to_contiguous(
    const char *__restrict__ src1_original, char *__restrict__ src1_contiguous,
    const mmid_row_mapping *__restrict__ row_mapping,
@@ -5643,6 +5732,7 @@ void ggml_backend_sycl_get_device_memory(int device, size_t * free, size_t * tot
    if (!res) {
        GGML_ABORT("[%s] failed to get device memory size", __func__);
    }
+    ggml_sycl_memtrace_report_device("device memory query", device, *free, *total);
 } catch (const sycl::exception & exc) {
    std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
    std::exit(1);
@@ -5890,6 +5980,17 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
            continue;
        }

+        // Batch consecutive independent same-shape F32 L2_NORM siblings (the GDN q/k
+        // norms) into one launch; sources are strided views of the fused qkv buffer, so
+        // the scan skips the interleaved view nodes instead of breaking on them.
+        if (node->op == GGML_OP_L2_NORM) {
+            const int l2_batch_skip = ggml_sycl_l2_norm_batch_fused(*sycl_ctx, cgraph, i);
+            if (l2_batch_skip > 0) {
+                i += l2_batch_skip;
+                continue;
+            }
+        }
+
        if (node->op == GGML_OP_MUL_MAT && ggml_sycl_mul_mat_glu_mmvq_fused(*sycl_ctx, cgraph, i)) {
            i += 2;
            continue;
@@ -6082,6 +6183,7 @@ static void ggml_backend_sycl_device_get_memory(ggml_backend_dev_t dev, size_t *
    if (!res) {
        GGML_ABORT("[%s] failed to get device memory size", __func__);
    }
+    ggml_sycl_memtrace_report_device("device memory query (dev)", ctx->device, *free, *total);
 }

 static enum ggml_backend_dev_type ggml_backend_sycl_device_get_type(ggml_backend_dev_t dev) {
@@ -0,0 +1,194 @@
+#include "memtrace.hpp"
+
+#include "common.hpp"
+#include "ggml-impl.h"
+
+#include <cstdio>
+#include <mutex>
+#include <unordered_map>
+
+constexpr size_t MIB = 1024 * 1024;
+
+static const char * mem_type_name(ggml_sycl_mem_type type) {
+    switch (type) {
+        case GGML_SYCL_MEM_BUFFER:   return "buffer";
+        case GGML_SYCL_MEM_POOL_LEG: return "pool_leg";
+        case GGML_SYCL_MEM_POOL_VMM: return "pool_vmm";
+        case GGML_SYCL_MEM_ASYNC:    return "async";
+        case GGML_SYCL_MEM_FATTN_KV: return "fattn_kv";
+        case GGML_SYCL_MEM_DIRECT:   return "direct";
+        default:                     GGML_ABORT("[%s] The type value %d is not supported\n", __func__, (int) type);
+    }
+}
+
+struct mem_tracker {
+    std::mutex mutex;
+    std::unordered_map<const void *, std::pair<ggml_sycl_mem_type, size_t>> live_by_ptr;
+    size_t live[GGML_SYCL_MEM_TYPE_COUNT] = {};
+    size_t peak[GGML_SYCL_MEM_TYPE_COUNT] = {};
+    size_t total_live = 0;
+    size_t total_peak = 0;
+    size_t last_logged_peak = 0;
+};
+
+static mem_tracker & get_tracker() {
+    static mem_tracker t;
+    return t;
+}
+
+static size_t step_bytes() {
+    const int mib = g_ggml_sycl_memtrace_step > 0 ? g_ggml_sycl_memtrace_step : 64;
+    return (size_t) mib * MIB;
+}
+
+static void report_sites_locked() {
+    mem_tracker & t = get_tracker();
+    for (int i = 0; i < GGML_SYCL_MEM_TYPE_COUNT; i++) {
+        if (t.peak[i] == 0) {
+            continue;
+        }
+        GGML_LOG_INFO(GGML_SYCL_MEMTRACE_TAG "   %-9s allocated %5zu MiB, peak %5zu MiB\n",
+                      mem_type_name((ggml_sycl_mem_type) i), t.live[i] / MIB, t.peak[i] / MIB);
+    }
+}
+
+static void report_locked(const char * tag) {
+    mem_tracker & t = get_tracker();
+
+    const size_t allocated = t.total_live / MIB;
+    const size_t buffers   = t.live[GGML_SYCL_MEM_BUFFER] / MIB;
+
+    GGML_LOG_INFO(GGML_SYCL_MEMTRACE_TAG " %s: allocated %5zu MiB (buffers %5zu + scratch %5zu),"
+                  " peak %5zu MiB\n",
+                  tag, allocated, buffers, allocated - buffers, t.total_peak / MIB);
+    report_sites_locked();
+}
+
+static void log_event_locked(const char * op, ggml_sycl_mem_type type, const void * ptr, size_t bytes) {
+    GGML_LOG_INFO(GGML_SYCL_MEMTRACE_TAG " allocated %5zu MiB  %-5s %-9s %9.3f MiB  ptr=%p\n",
+                  get_tracker().total_live / MIB, op, mem_type_name(type),
+                  (double) bytes / MIB, ptr);
+}
+
+bool ggml_sycl_memtrace_enabled() {
+    return g_ggml_sycl_memtrace > 0;
+}
+
+void ggml_sycl_memtrace_add(ggml_sycl_mem_type type, const void * ptr, size_t bytes) {
+    if (!ggml_sycl_memtrace_enabled()) {
+        return;
+    }
+    GGML_ASSERT(ptr != nullptr);
+    GGML_ASSERT(bytes != 0);
+
+    mem_tracker & t = get_tracker();
+    std::lock_guard<std::mutex> lock(t.mutex);
+
+    auto it = t.live_by_ptr.find(ptr);
+    if (it != t.live_by_ptr.end()) {
+        t.live[it->second.first] -= it->second.second;
+        t.total_live -= it->second.second;
+    }
+
+    t.live_by_ptr[ptr] = { type, bytes };
+    t.live[type] += bytes;
+    t.total_live += bytes;
+
+    if (t.live[type] > t.peak[type]) {
+        t.peak[type] = t.live[type];
+    }
+    if (t.total_live > t.total_peak) {
+        t.total_peak = t.total_live;
+    }
+
+    if (g_ggml_sycl_memtrace >= 2) {
+        log_event_locked("alloc", type, ptr, bytes);
+    }
+
+    static const size_t step = step_bytes();
+    if (t.total_peak >= t.last_logged_peak + step) {
+        t.last_logged_peak = t.total_peak;
+        char tag[96];
+        std::snprintf(tag, sizeof(tag), "peak grew (+%zu MiB from %s)", bytes / MIB,
+                      mem_type_name(type));
+        report_locked(tag);
+    }
+}
+
+void ggml_sycl_memtrace_del(const void * ptr) {
+    if (!ggml_sycl_memtrace_enabled() || ptr == nullptr) {
+        return;
+    }
+    mem_tracker & t = get_tracker();
+    std::lock_guard<std::mutex> lock(t.mutex);
+
+    auto it = t.live_by_ptr.find(ptr);
+    if (it == t.live_by_ptr.end()) {
+        return;
+    }
+    const ggml_sycl_mem_type type = it->second.first;
+    const size_t bytes = it->second.second;
+    t.live[type] -= bytes;
+    t.total_live -= bytes;
+    t.live_by_ptr.erase(it);
+
+    if (g_ggml_sycl_memtrace >= 2) {
+        log_event_locked("free", type, ptr, bytes);
+    }
+}
+
+void ggml_sycl_memtrace_fail(ggml_sycl_mem_type type, size_t bytes) {
+    GGML_LOG_ERROR(GGML_SYCL_MEMTRACE_TAG " alloc FAILED: %9.3f MiB %s\n",
+                   (double) bytes / MIB, mem_type_name(type));
+    if (!ggml_sycl_memtrace_enabled()) {
+        return;
+    }
+    mem_tracker & t = get_tracker();
+    std::lock_guard<std::mutex> lock(t.mutex);
+    report_locked("at allocation failure");
+}
+
+void ggml_sycl_memtrace_report(const char * tag) {
+    if (!ggml_sycl_memtrace_enabled()) {
+        return;
+    }
+    mem_tracker & t = get_tracker();
+    std::lock_guard<std::mutex> lock(t.mutex);
+    report_locked(tag);
+}
+
+static bool device_memory_is_dedicated(int device) {
+    if (device < 0 || device >= ggml_sycl_info().device_count) {
+        return false;
+    }
+    const sycl_device_info & info = ggml_sycl_info().devices[device];
+    return info.l0_device_type_valid && info.l0_discrete_gpu;
+}
+
+void ggml_sycl_memtrace_report_device(const char * tag, int device, size_t dev_free, size_t dev_total) {
+    if (!ggml_sycl_memtrace_enabled()) {
+        return;
+    }
+    mem_tracker & t = get_tracker();
+    std::lock_guard<std::mutex> lock(t.mutex);
+
+    const size_t in_use    = dev_total > dev_free ? dev_total - dev_free : 0;
+    const size_t total     = dev_total / MIB;
+    const size_t freed     = dev_free / MIB;
+    const size_t allocated = t.total_live / MIB;
+    const size_t buffers   = t.live[GGML_SYCL_MEM_BUFFER] / MIB;
+    const size_t peak      = t.total_peak / MIB;
+
+    if (in_use >= t.total_live && device_memory_is_dedicated(device) && total >= freed + allocated) {
+        GGML_LOG_INFO(GGML_SYCL_MEMTRACE_TAG " %s: total %5zu MiB = free %5zu + allocated %5zu"
+                      " (buffers %5zu + scratch %5zu) + other %5zu, peak %5zu MiB\n",
+                      tag, total, freed, allocated, buffers, allocated - buffers,
+                      total - freed - allocated, peak);
+    } else {
+        GGML_LOG_INFO(GGML_SYCL_MEMTRACE_TAG " %s: total %5zu MiB, free %5zu, in use %5zu;"
+                      " allocated %5zu (buffers %5zu + scratch %5zu), peak %5zu MiB\n",
+                      tag, total, freed, in_use / MIB, allocated, buffers,
+                      allocated - buffers, peak);
+    }
+    report_sites_locked();
+}
@@ -0,0 +1,28 @@
+#ifndef GGML_SYCL_MEMTRACE_HPP
+#define GGML_SYCL_MEMTRACE_HPP
+
+#include <cstddef>
+
+#define GGML_SYCL_MEMTRACE_TAG "[SYCL-MEMTRACE]"
+
+enum ggml_sycl_mem_type {
+    GGML_SYCL_MEM_BUFFER = 0,
+    GGML_SYCL_MEM_POOL_LEG,
+    GGML_SYCL_MEM_POOL_VMM,
+    GGML_SYCL_MEM_ASYNC,
+    GGML_SYCL_MEM_FATTN_KV,
+    GGML_SYCL_MEM_DIRECT,
+
+    GGML_SYCL_MEM_TYPE_COUNT,
+};
+
+bool ggml_sycl_memtrace_enabled();
+
+void ggml_sycl_memtrace_add(ggml_sycl_mem_type type, const void * ptr, size_t bytes);
+void ggml_sycl_memtrace_del(const void * ptr);
+
+void ggml_sycl_memtrace_report(const char * tag);
+void ggml_sycl_memtrace_report_device(const char * tag, int device, size_t dev_free, size_t dev_total);
+void ggml_sycl_memtrace_fail(ggml_sycl_mem_type type, size_t bytes);
+
+#endif  // GGML_SYCL_MEMTRACE_HPP
@@ -543,6 +543,62 @@ static void l2_norm_f32_sycl(const float *   x,
    }
 }

+// Batched L2 norm: N independent same-shape F32 tensors in one launch; the tensor
+// index is folded into grid dim0 and each row's reduction is identical to the
+// single-tensor kernel, so the result is bit-exact.
+struct l2_batch_ptrs {
+    const float * src[GGML_SYCL_L2_BATCH_MAX];
+    float *       dst[GGML_SYCL_L2_BATCH_MAX];
+};
+
+// One stride set shared by the whole batch: the caller only groups tensors whose nb[]
+// all match, so per-tensor state stays two pointers.
+struct l2_batch_strides {
+    int     ne1, ne2;
+    int64_t ss0, ss1, ss2, ss3;
+    int64_t ds0, ds1, ds2, ds3;
+};
+
+template <int warp_size>
+static void l2_norm_f32_batch(l2_batch_ptrs p, l2_batch_strides st, const int ncols, const float eps,
+                              const sycl::nd_item<3> & item_ct1) {
+    const int t   = item_ct1.get_group(0);  // tensor index
+    const int r   = item_ct1.get_group(2);  // flattened row over ne1*ne2*ne3
+    const int tid = item_ct1.get_local_id(2);
+
+    const int i1 = r % st.ne1;
+    const int i2 = (r / st.ne1) % st.ne2;
+    const int i3 = r / (st.ne1 * st.ne2);
+
+    const float * x   = p.src[t] + i3 * st.ss3 + i2 * st.ss2 + i1 * st.ss1;
+    float *       dst = p.dst[t] + i3 * st.ds3 + i2 * st.ds2 + i1 * st.ds1;
+
+    float tmp = 0.0f;
+    for (int col = tid; col < ncols; col += warp_size) {
+        const float xi = x[col * st.ss0];
+        tmp += xi * xi;
+    }
+    tmp = block_reduce<block_reduce_method::SUM, warp_size>(tmp, (float *) nullptr, warp_size);
+    const float scale = sycl::rsqrt(sycl::fmax(tmp, eps * eps));
+    for (int col = tid; col < ncols; col += warp_size) {
+        dst[col * st.ds0] = scale * x[col * st.ss0];
+    }
+}
+
+template <int warp_size>
+static void l2_norm_f32_batch_sycl(l2_batch_ptrs p, l2_batch_strides st, const int n_tensors,
+                                   const int ncols, const int nrows_total, const float eps,
+                                   queue_ptr stream) {
+    const dpct::dim3 blocks_num(nrows_total, 1, n_tensors);
+    const dpct::dim3 block_dims(warp_size, 1, 1);
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(blocks_num * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(warp_size)]] {
+                l2_norm_f32_batch<warp_size>(p, st, ncols, eps, item_ct1);
+            });
+    });
+}
+
 void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
    const ggml_tensor * src0 = dst->src[0];

@@ -961,3 +1017,30 @@ void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
    l2_norm_f32_sycl<WARP_SIZE>(src0_d, dst_d, ne00, ne01, ne02, ne03,
            ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, stream, ctx.device);
 }
+
+// nodes[0..count) are independent, same-shape, same-eps, same-nb L2_NORM ops validated
+// by the caller; requires ncols < 1024 (the warp reduction path).
+void ggml_sycl_l2_norm_batch(ggml_backend_sycl_context & ctx, ggml_tensor ** nodes, int count) {
+    const ggml_tensor * s0 = nodes[0]->src[0];
+    const int ncols       = (int) s0->ne[0];
+    const int nrows_total = (int) ggml_nrows(s0);
+    float eps;
+    memcpy(&eps, nodes[0]->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
+
+    l2_batch_ptrs p{};
+    for (int t = 0; t < count; ++t) {
+        p.src[t] = (const float *) nodes[t]->src[0]->data;
+        p.dst[t] = (float *) nodes[t]->data;
+    }
+
+    const ggml_tensor * d0 = nodes[0];
+    const size_t        ts = ggml_type_size(GGML_TYPE_F32);
+    l2_batch_strides    st{};
+    st.ne1 = (int) s0->ne[1];
+    st.ne2 = (int) s0->ne[2];
+    st.ss0 = s0->nb[0] / ts; st.ss1 = s0->nb[1] / ts; st.ss2 = s0->nb[2] / ts; st.ss3 = s0->nb[3] / ts;
+    st.ds0 = d0->nb[0] / ts; st.ds1 = d0->nb[1] / ts; st.ds2 = d0->nb[2] / ts; st.ds3 = d0->nb[3] / ts;
+
+    l2_norm_f32_batch_sycl<WARP_SIZE>(p, st, count, ncols, nrows_total, eps, ctx.stream());
+}
@@ -29,4 +29,7 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);

 void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);

+#define GGML_SYCL_L2_BATCH_MAX 8
+void ggml_sycl_l2_norm_batch(ggml_backend_sycl_context & ctx, ggml_tensor ** nodes, int count);
+
 #endif // GGML_SYCL_NORM_HPP
--- a/Show More
+++ b/Show More