Revert "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3 (#24546 )" (#28551 )

This reverts commit 0c963452ea. Assisted-by: Claude Fable 5.1 Claude-Session: https://claude.ai/code/session_01Q7rfnkjzgnfhvJsdeXhdoH
nix : update deprecated expressions (#28145 )
2026-09-07 16:37:57 +02:00 · 2026-09-07 16:28:19 +02:00 · 2026-09-07 15:59:45 +02:00 · 2026-09-07 15:55:14 +02:00 · 2026-09-07 15:50:46 +02:00 · 2026-09-07 15:24:14 +02:00
166 changed files with 8764 additions and 1431 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: |
@@ -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

@@ -3901,6 +3901,14 @@ 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)",
@@ -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
@@ -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)
@@ -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. |
@@ -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; };
@@ -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;
-                        }
-                    }
                }
            }

@@ -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)
@@ -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
@@ -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;
        }

@@ -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) {
@@ -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 } },
@@ -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;
    }
@@ -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
@@ -608,6 +608,21 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif

+#if defined(DATA_A_TQ1_0)
+float tq1_0_val(uint ib, uint e, uint a_offset) {
+    const uint bidx = tq1_0_byte_of(e);
+    const uint qbyte = uint(bidx < 48u ? data_a[a_offset + ib].qs[bidx]
+                                       : data_a[a_offset + ib].qh[bidx - 48u]);
+    return float(tq1_0_trit(qbyte, tq1_0_digit_of(e))) - 1.0;
+}
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(tq1_0_val(ib, iqs, a_offset), tq1_0_val(ib, iqs + 1u, a_offset));
+}
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
+#endif
+
 #if defined(DATA_A_TQ2_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
    // elem e -> byte qs[(e/128)*32 + e%32], bits 2*((e%128)/32); w = q - 1 (d applied via get_dm)
@@ -247,6 +247,19 @@ f16vec4 dequantFuncQ8_0_v(const in decodeBufQ8_0 bl, const in uint blockCoords[2
    return f16vec4(vec4(qi) * vec4(float(d)));
 }

+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufTQ1_0 {
+   block_tq1_0 block;
+};
+
+float16_t dequantFuncTQ1_0(const in decodeBufTQ1_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint e = coordInBlock[1];
+    const uint bidx = tq1_0_byte_of(e);
+    const uint qbyte = uint(bidx < 48u ? bl.block.qs[bidx] : bl.block.qh[bidx - 48u]);
+    const uint xi = tq1_0_trit(qbyte, tq1_0_digit_of(e));
+    return bl.block.d * (float16_t(int(xi)) - float16_t(1.0));
+}
+
 layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufTQ2_0 {
   block_tq2_0 block;
 };
@@ -1406,6 +1419,8 @@ f16vec4 dequantFuncNVFP4_v(const in decodeBufNVFP4 bl, const in uint blockCoords
 #elif defined(DATA_A_Q8_0)
 #define dequantFuncA dequantFuncQ8_0
 #define dequantFuncA_v dequantFuncQ8_0_v
+#elif defined(DATA_A_TQ1_0)
+#define dequantFuncA dequantFuncTQ1_0
 #elif defined(DATA_A_TQ2_0)
 #define dequantFuncA dequantFuncTQ2_0
 #define dequantFuncA_v dequantFuncTQ2_0_v
@@ -0,0 +1,28 @@
+#version 450
+
+#include "dequant_head.glsl"
+
+layout (local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_tq1_0 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x * 4;
+
+    if (i >= p.nel) {
+        return;
+    }
+
+    const uint ib = i / QUANT_K_TQ1_0;
+    const float d = float(data_a[ib].d);
+
+    [[unroll]] for (uint j = 0; j < 4 && (i + j) < p.nel; ++j) {
+        const uint e = (i + j) % QUANT_K_TQ1_0;
+        const uint bidx = tq1_0_byte_of(e);
+        const uint qbyte = uint(bidx < 48u ? data_a[ib].qs[bidx]
+                                           : data_a[ib].qh[bidx - 48u]);
+        const uint xi = tq1_0_trit(qbyte, tq1_0_digit_of(e));
+        data_b[i + j] = D_TYPE(d * (float(xi) - 1.0f));
+    }
+}
@@ -0,0 +1,90 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_shuffle : require
+
+// 16 lanes per token, indexed idst + hc*isrc: idst in bits 0..1, isrc in bits 2..3,
+// so subgroupShuffleXor by 1|2 reduces a row and by 4|8 a column.
+
+layout(constant_id = 0) const uint SUBGROUP_SIZE = 32;
+
+layout(local_size_x_id = 0, local_size_y = 4, local_size_z = 1) in;
+
+layout(push_constant) uniform parameter
+{
+    uint n_tokens;
+
+    uint nbm0; uint nbm1;   // mixes
+    uint nbs0;              // scale
+    uint nbb0;              // base
+    uint nbd0; uint nbd1; uint nbd2;   // dst
+
+    uint m_offset;
+    uint s_offset;
+    uint b_offset;
+    uint d_offset;
+
+    float eps;
+    uint n_iter;
+};
+
+layout(binding = 0, std430) readonly buffer M { float data_m[]; };
+layout(binding = 1, std430) readonly buffer S { float data_s[]; };
+layout(binding = 2, std430) readonly buffer B { float data_b[]; };
+layout(binding = 3, std430) writeonly buffer D { float data_d[]; };
+
+const uint hc          = 4;
+const uint comb_offset = 2 * hc;
+
+const uint TOKENS_PER_SUBGROUP = SUBGROUP_SIZE / 16;
+
+void main() {
+    const uint lane = gl_SubgroupInvocationID;
+    const uint blk  = lane >> 4;    // which 16-lane block, i.e. which token
+    const uint idx  = lane & 15;    // idst + hc*isrc
+
+    const uint sg = gl_WorkGroupID.x * gl_WorkGroupSize.y + gl_SubgroupID;
+    const uint it = sg * TOKENS_PER_SUBGROUP + blk;
+
+    // no early return, the shuffles need every lane; out-of-range blocks compute a discarded value
+    const bool in_range = it < n_tokens;
+
+    const float scale_comb = data_s[s_offset + 2 * nbs0];
+
+    float v = 0.0f;
+    if (in_range) {
+        v = data_m[m_offset + (comb_offset + idx) * nbm0 + it * nbm1] * scale_comb
+          + data_b[b_offset + (comb_offset + idx) * nbb0];
+    }
+
+    // Softmax across destinations: the four lanes sharing an isrc.
+    float vmax = max(v, subgroupShuffleXor(v, 1));
+    vmax = max(vmax, subgroupShuffleXor(vmax, 2));
+    v = exp(v - vmax);
+
+    float sum = v + subgroupShuffleXor(v, 1);
+    sum += subgroupShuffleXor(sum, 2);
+    v = v / sum + eps;
+
+    // Normalize columns: equal destination indices are four lanes apart.
+    sum = v + subgroupShuffleXor(v, 4);
+    sum += subgroupShuffleXor(sum, 8);
+    v /= sum + eps;
+
+    for (uint i = 1; i < n_iter; ++i) {
+        sum = v + subgroupShuffleXor(v, 1);
+        sum += subgroupShuffleXor(sum, 2);
+        v /= sum + eps;
+
+        sum = v + subgroupShuffleXor(v, 4);
+        sum += subgroupShuffleXor(sum, 8);
+        v /= sum + eps;
+    }
+
+    if (in_range) {
+        const uint idst = idx & 3;
+        const uint isrc = idx >> 2;
+        data_d[d_offset + idst * nbd0 + isrc * nbd1 + it * nbd2] = v;
+    }
+}
@@ -0,0 +1,83 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+// Fan one stream back out to hc streams and add the combination-weighted
+// residuals:
+//
+//   dst[i0, idst, it] = x[i0, it]*post[idst, it]
+//                     + sum_isrc residual[i0, isrc, it]*comb[idst, isrc, it]
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 256;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform parameter
+{
+    uint n_embd;
+    uint n_tokens;
+
+    uint nbx0; uint nbx1;              // x
+    uint nbr0; uint nbr1; uint nbr2;   // residual
+    uint nbp0; uint nbp1;              // post
+    uint nbc0; uint nbc1; uint nbc2;   // comb
+    uint nbd0; uint nbd1; uint nbd2;   // dst
+
+    uint x_offset;
+    uint r_offset;
+    uint p_offset;
+    uint c_offset;
+    uint d_offset;
+};
+
+layout(binding = 0, std430) readonly buffer X { float data_x[]; };
+layout(binding = 1, std430) readonly buffer R { float data_r[]; };
+layout(binding = 2, std430) readonly buffer P { float data_p[]; };
+layout(binding = 3, std430) readonly buffer C { float data_c[]; };
+layout(binding = 4, std430) writeonly buffer D { float data_d[]; };
+
+const uint hc = 4;
+
+shared float post_s[hc];
+shared float comb_s[hc * hc];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint it  = gl_WorkGroupID.y;
+
+    if (tid < hc) {
+        post_s[tid] = data_p[p_offset + tid * nbp0 + it * nbp1];
+    }
+    if (tid < hc * hc) {
+        const uint idst = tid & 3;
+        const uint isrc = tid >> 2;
+        comb_s[tid] = data_c[c_offset + idst * nbc0 + isrc * nbc1 + it * nbc2];
+    }
+    barrier();
+
+    // After the barrier, so every invocation reaches it.
+    const uint i0 = gl_WorkGroupID.x * BLOCK_SIZE + tid;
+    if (i0 >= n_embd) {
+        return;
+    }
+
+    const float xv = data_x[x_offset + i0 * nbx0 + it * nbx1];
+
+    const uint rb = r_offset + i0 * nbr0 + it * nbr2;
+
+    float r[hc];
+    [[unroll]]
+    for (uint isrc = 0; isrc < hc; ++isrc) {
+        r[isrc] = data_r[rb + isrc * nbr1];
+    }
+
+    [[unroll]]
+    for (uint idst = 0; idst < hc; ++idst) {
+        float result = xv * post_s[idst];
+        [[unroll]]
+        for (uint isrc = 0; isrc < hc; ++isrc) {
+            result = fma(r[isrc], comb_s[idst + hc * isrc], result);
+        }
+        data_d[d_offset + i0 * nbd0 + idst * nbd1 + it * nbd2] = result;
+    }
+}
@@ -0,0 +1,59 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+// Collapse the hc residual streams of a token into one, weighted per stream:
+//
+//   dst[i0, it] = sum_ih x[i0, ih, it] * weights[ih, it]
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 256;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform parameter
+{
+    uint n_embd;
+    uint n_tokens;
+
+    uint nbx0; uint nbx1; uint nbx2;   // x
+    uint nbw0; uint nbw1;              // weights
+    uint nbd0; uint nbd1;              // dst
+
+    uint x_offset;
+    uint w_offset;
+    uint d_offset;
+};
+
+layout(binding = 0, std430) readonly buffer X { float data_x[]; };
+layout(binding = 1, std430) readonly buffer W { float data_w[]; };
+layout(binding = 2, std430) writeonly buffer D { float data_d[]; };
+
+const uint hc = 4;
+
+shared float w[hc];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint it  = gl_WorkGroupID.y;
+
+    if (tid < hc) {
+        w[tid] = data_w[w_offset + tid * nbw0 + it * nbw1];
+    }
+    barrier();
+
+    // After the barrier, so every invocation reaches it.
+    const uint i0 = gl_WorkGroupID.x * BLOCK_SIZE + tid;
+    if (i0 >= n_embd) {
+        return;
+    }
+
+    const uint xb = x_offset + i0 * nbx0 + it * nbx2;
+
+    float result = 0.0f;
+    [[unroll]]
+    for (uint ih = 0; ih < hc; ++ih) {
+        result = fma(data_x[xb + ih * nbx1], w[ih], result);
+    }
+
+    data_d[d_offset + i0 * nbd0 + it * nbd1] = result;
+}
@@ -27,10 +27,10 @@ void main() {
            const uint i11 = gid_z / p.ne12;
            const uint i12 = gid_z % p.ne12;

-            const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+            const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12];

-            const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
-            const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+            const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+            const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;

            const uint ib = a_offset + i00/QUANT_K; // block index
            const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index
@@ -0,0 +1,85 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#include "mul_mat_vec_base.glsl"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+// Walks the packed bytes directly (byte m, digit t) rather than via
+// tq1_0_byte_of()/tq1_0_digit_of(): one byte per thread, expanded in place.
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint tid = gl_LocalInvocationID.x;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    for (uint nrow = 0; nrow < num_rows; ++nrow) {
+        const uint ib0 = a_offset + (first_row + nrow) * num_blocks_per_row;
+        for (uint jcol = 0; jcol < NUM_COLS; ++jcol) {
+            const uint b_base = (jcol * p.batch_stride_b);
+            for (uint i = tid/8; i < num_blocks_per_row; i += gl_WorkGroupSize.x/8) {
+                const FLOAT_TYPE d = float(data_a[ib0 + i].d);
+
+                // First qs chunk: 32 bytes (5*32 elements)
+                [[unroll]] for (uint m = tid%8; m < 32; m += 8) {
+                    const uint q_byte = uint(data_a[ib0 + i].qs[m]);
+                    [[unroll]] for (uint t = 0; t < 5; ++t) {
+                        const uint xi = tq1_0_trit(q_byte, t);
+                        const FLOAT_TYPE dequant_val = FLOAT_TYPE(d * (float(xi) - 1.0f));
+                        const uint elem = t * 32u + m;
+                        const uint b_idx = i * QUANT_K + elem;
+                        temp[jcol][nrow] += dequant_val * FLOAT_TYPE(data_b[b_base + b_offset + b_idx]);
+                    }
+                }
+
+                // Second qs chunk: 16 bytes (5*16 elements)
+                [[unroll]] for (uint m = tid%8; m < 16; m += 8) {
+                    const uint q_byte = uint(data_a[ib0 + i].qs[32u + m]);
+                    [[unroll]] for (uint t = 0; t < 5; ++t) {
+                        const uint xi = tq1_0_trit(q_byte, t);
+                        const FLOAT_TYPE dequant_val = FLOAT_TYPE(d * (float(xi) - 1.0f));
+                        const uint elem = 160u + t * 16u + m;
+                        const uint b_idx = i * QUANT_K + elem;
+                        temp[jcol][nrow] += dequant_val * FLOAT_TYPE(data_b[b_base + b_offset + b_idx]);
+                    }
+                }
+
+                // qh bytes: 4 bytes (4*4 elements)
+                [[unroll]] for (uint j = tid%8; j < 4; j += 8) {
+                    const uint qh_byte = uint(data_a[ib0 + i].qh[j]);
+                    [[unroll]] for (uint t = 0; t < 4; ++t) {
+                        const uint xi = tq1_0_trit(qh_byte, t);
+                        const FLOAT_TYPE dequant_val = FLOAT_TYPE(d * (float(xi) - 1.0f));
+                        const uint elem = 240u + t * 4u + j;
+                        const uint b_idx = i * QUANT_K + elem;
+                        temp[jcol][nrow] += dequant_val * FLOAT_TYPE(data_b[b_base + b_offset + b_idx]);
+                    }
+                }
+            }
+        }
+    }
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
@@ -197,6 +197,24 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const uint k_pair = row * LOAD_VEC_A / 2;
            store_a(col, k_pair,     FLOAT_TYPEV2(v.xy));
            store_a(col, k_pair + 1, FLOAT_TYPEV2(v.zw));
+#elif defined(DATA_A_TQ1_0)
+            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
+
+            const uint ib  = idx / 128;               // 2 values per idx
+            const uint iqs = (idx % 128) * 2;         // element 0,2,4..254
+
+            const float d = float(data_a[ib].d);
+            vec2 v;
+            for (uint kk = 0u; kk < 2u; ++kk) {
+                const uint e = iqs + kk;
+                const uint bidx = tq1_0_byte_of(e);
+                const uint qbyte = uint(bidx < 48u ? data_a[ib].qs[bidx]
+                                                   : data_a[ib].qh[bidx - 48u]);
+                v[kk] = d * (float(tq1_0_trit(qbyte, tq1_0_digit_of(e))) - 1.0);
+            }
+
+            const uint k_pair = row * LOAD_VEC_A / 2;
+            store_a(col, k_pair, FLOAT_TYPEV2(v.xy));
 #elif defined(DATA_A_TQ2_0)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;

@@ -27,12 +27,24 @@ layout (binding = 6) readonly buffer R_I {uvec2 rope_data_i[];}; // indices for
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24

+#elif RMS_NORM_ADD_FUSION
+
+layout (binding = 3) readonly buffer C {float data_c[];};
+layout (binding = 4) readonly buffer E {float data_e[];};
+
+#elif RMS_NORM_SET_ROWS_FUSION
+
+layout (binding = 3) readonly buffer I {uvec2 data_i[];};
+
 #endif

 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512

 layout (constant_id = 1) const bool do_multiply = false;
+#if RMS_NORM_ADD_FUSION
+layout (constant_id = 2) const bool do_post_multiply = false;
+#endif

 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;

@@ -57,6 +69,8 @@ void rms_norm(uint num_iters) {
 #if RMS_NORM_ROPE_FUSION
    // Per-row offset in shared memory
    uint32_t d_offset = 0;
+#elif RMS_NORM_SET_ROWS_FUSION
+    uint32_t d_offset = data_i[channel].x*p.nb21 + row*ncols + get_doffset();
 #else
    uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
 #endif
@@ -91,14 +105,28 @@ void rms_norm(uint num_iters) {
                if (col >= ncols) {
                    continue;
                }
-                data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
+                FLOAT_TYPE value = scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]);
+#if RMS_NORM_ADD_FUSION
+                value += FLOAT_TYPE(data_c[d_offset + col]);
+                if (do_post_multiply) {
+                    value *= FLOAT_TYPE(data_e[0]);
+                }
+#endif
+                data_d[d_offset + col] = D_TYPE(value);
            }
        } else {
            [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
                if (col >= ncols) {
                    continue;
                }
-                data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+                FLOAT_TYPE value = scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]);
+#if RMS_NORM_ADD_FUSION
+                value += FLOAT_TYPE(data_c[d_offset + col]);
+                if (do_post_multiply) {
+                    value *= FLOAT_TYPE(data_e[0]);
+                }
+#endif
+                data_d[d_offset + col] = D_TYPE(value);
            }
        }
    } else {
@@ -10,11 +10,19 @@
 #define BLOCK_SIZE 128

 layout (constant_id = 1) const bool do_multiply = false;
+#if RMS_NORM_ADD_FUSION
+layout (constant_id = 2) const bool do_post_multiply = false;
+#endif

 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;

 layout (binding = 3, std430) readonly buffer PartialsBuf {float partial_sums[];};

+#if RMS_NORM_ADD_FUSION
+layout (binding = 4) readonly buffer C {float data_c[];};
+layout (binding = 5) readonly buffer E {float data_e[];};
+#endif
+
 shared FLOAT_TYPE sumsh[BLOCK_SIZE];

 void main() {
@@ -55,9 +63,23 @@ void main() {

    if (do_multiply) {
        if (ncols > p.ne10) {
-            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]));
+            FLOAT_TYPE value = scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + fastmod(col, p.ne10)]);
+#if RMS_NORM_ADD_FUSION
+            value += FLOAT_TYPE(data_c[d_offset + col]);
+            if (do_post_multiply) {
+                value *= FLOAT_TYPE(data_e[0]);
+            }
+#endif
+            data_d[d_offset + col] = D_TYPE(value);
        } else {
-            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+            FLOAT_TYPE value = scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]);
+#if RMS_NORM_ADD_FUSION
+            value += FLOAT_TYPE(data_c[d_offset + col]);
+            if (do_post_multiply) {
+                value *= FLOAT_TYPE(data_e[0]);
+            }
+#endif
+            data_d[d_offset + col] = D_TYPE(value);
        }
    } else {
        data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
@@ -303,6 +303,41 @@ struct block_q2_K_packed32
 #define DATA_A_QUANT_K
 #endif

+#define QUANT_K_TQ1_0 256
+
+// TQ1_0: base-3 packed trits, 5 per byte in `qs` (48B) and 4 in `qh` (4B).
+struct block_tq1_0
+{
+    uint8_t qs[(QUANT_K_TQ1_0 - 4 * QUANT_K_TQ1_0 / 64) / 5];
+    uint8_t qh[QUANT_K_TQ1_0 / 64];
+    float16_t d;
+};
+
+// Element e in [0,255] -> its packed byte (0..47 qs, 48..51 qh) and digit.
+uint tq1_0_byte_of(uint e) {
+    return e < 160u ? (e % 32u)
+         : e < 240u ? 32u + ((e - 160u) % 16u)
+         : 48u + ((e - 240u) % 4u);
+}
+uint tq1_0_digit_of(uint e) {
+    return e < 160u ? (e / 32u)
+         : e < 240u ? ((e - 160u) / 16u)
+         : ((e - 240u) / 4u);
+}
+// The 8-bit truncation below is part of the format, not an optimisation:
+// the C reference does `uint8_t q = qs[..] * pow3[n]`.
+uint tq1_0_trit(uint qbyte, uint t) {
+    const uint POW3_PACKED = (1u << 28) | (3u << 21) | (9u << 14) | (27u << 7) | 81u;
+    return ((((qbyte * ((POW3_PACKED >> (7u * (4u - t))) & 0x7Fu)) & 255u) * 3u) >> 8);
+}
+
+#if defined(DATA_A_TQ1_0)
+#define QUANT_K QUANT_K_TQ1_0
+#define QUANT_R 1
+#define A_TYPE block_tq1_0
+#define DATA_A_QUANT_K
+#endif
+
 #define QUANT_K_TQ2_0 256

 // ternary (BitNet): 2-bit codes, w = (q - 1) * d; qs layout matches q2_K's
@@ -72,6 +72,7 @@ const std::vector<std::string> type_names = {
    "iq4_nl",
    "mxfp4",
    "nvfp4",
+    "tq1_0",
    "tq2_0",
    "bf16",
 };
@@ -734,7 +735,7 @@ void process_shaders() {
    for (const auto& tname : type_names) {
        // mul mat vec
        std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_") || tname == "tq2_0") ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
+        std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_") || tname == "tq2_0" || tname == "tq1_0") ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";

        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPEV2", "vec2"}, {"B_TYPEV4", "vec4"}, {"D_TYPE", "float"}}));
        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPEV2", "f16vec2"}, {"B_TYPEV4", "f16vec4"}, {"D_TYPE", "float"}}));
@@ -805,6 +806,10 @@ void process_shaders() {
    string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("rms_norm_mul_add_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"RMS_NORM_ADD_FUSION", "1"}}));
+    string_to_spv("rms_norm_mul_add_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"RMS_NORM_ADD_FUSION", "1"}}));
+    string_to_spv("rms_norm_set_rows_f32_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"RMS_NORM_SET_ROWS_FUSION", "1"}}));
+    string_to_spv("rms_norm_set_rows_f32_f16", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RMS_NORM_SET_ROWS_FUSION", "1"}}));
    string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
    string_to_spv("rms_norm_mul_rope_f32_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float"}, {"RMS_NORM_ROPE_FUSION", "1"}}));
    string_to_spv("rms_norm_mul_rope_f32_f16", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RMS_NORM_ROPE_FUSION", "1"}}));
@@ -1037,6 +1042,9 @@ void process_shaders() {
    string_to_spv("fwht_f32", "fwht.comp", {});
    string_to_spv("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}});
    string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
+    string_to_spv("dsv4_hc_comb_f32", "dsv4_hc_comb.comp", {});
+    string_to_spv("dsv4_hc_pre_f32",  "dsv4_hc_pre.comp",  {});
+    string_to_spv("dsv4_hc_post_f32", "dsv4_hc_post.comp", {});
    string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    string_to_spv("cumsum_multipass1_f32", "cumsum_multipass1.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
    string_to_spv("cumsum_multipass2_f32", "cumsum_multipass2.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
@@ -4324,12 +4324,22 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                           src0->type == GGML_TYPE_F32 && (src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32));
            break;
        case GGML_OP_GET_ROWS:
-            if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_webgpu_supported_qtype(src0->type)) {
-                supports_op = (op->type == GGML_TYPE_F32);
-            } else if (src0->type == GGML_TYPE_I32) {
-                supports_op = op->type == GGML_TYPE_I32;
+            {
+                const size_t storage_alignment =
+                    ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment;
+                const size_t src_address_unit =
+                    src0->type == GGML_TYPE_F32 && op->ne[0] % 4 == 0 ? 4 * sizeof(float) : ggml_type_size(src0->type);
+                if (ggml_webgpu_tensor_misalignment(src0, storage_alignment) % src_address_unit != 0) {
+                    break;
+                }
+                if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 ||
+                    ggml_webgpu_supported_qtype(src0->type)) {
+                    supports_op = (op->type == GGML_TYPE_F32);
+                } else if (src0->type == GGML_TYPE_I32) {
+                    supports_op = op->type == GGML_TYPE_I32;
+                }
+                break;
            }
-            break;
        case GGML_OP_MUL_MAT:
            {
                switch (src1->type) {
@@ -215,6 +215,7 @@ class Keys:
        KV_LORA_RANK_SWA             = "{arch}.attention.kv_lora_rank_swa"
        SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
        SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
+        RECURRENT_LAYERS             = "{arch}.attention.recurrent_layers"
        TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
        ROPE_PATTERN                 = "{arch}.attention.rope_pattern"

@@ -230,6 +231,8 @@ class Keys:
        COUNT                = "{arch}.hyper_connection.count"
        SINKHORN_ITERATIONS  = "{arch}.hyper_connection.sinkhorn_iterations"
        EPSILON              = "{arch}.hyper_connection.epsilon"
+        # scale of the post gate (DeepSeek-V4 hardcodes 2.0)
+        MAGNITUDE            = "{arch}.hyper_connection.magnitude"
        # absent means the mix projection is full rank (DeepSeek-V4 behaviour)
        LOW_RANK             = "{arch}.hyper_connection.low_rank"

@@ -592,6 +595,7 @@ class MODEL_ARCH(IntEnum):
    HUNYUAN_DENSE    = auto()
    HUNYUAN_VL       = auto()
    HY_V3            = auto()
+    HY_V4            = auto()
    SMOLLM3          = auto()
    GPT_OSS          = auto()
    LFM2             = auto()
@@ -616,6 +620,7 @@ class MODEL_ARCH(IntEnum):
    PADDLEOCR        = auto()
    MIMO2            = auto()
    STEP35           = auto()
+    SPARK2_5           = auto()
    LLAMA_EMBED      = auto()
    MAINCODER        = auto()
    KIMI_LINEAR      = auto()
@@ -1345,6 +1350,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.HUNYUAN_DENSE:    "hunyuan-dense",
    MODEL_ARCH.HUNYUAN_VL:       "hunyuan_vl",
    MODEL_ARCH.HY_V3:            "hy_v3",
+    MODEL_ARCH.HY_V4:            "hy_v4",
    MODEL_ARCH.SMOLLM3:          "smollm3",
    MODEL_ARCH.GPT_OSS:          "gpt-oss",
    MODEL_ARCH.LFM2:             "lfm2",
@@ -1369,6 +1375,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.PADDLEOCR:        "paddleocr",
    MODEL_ARCH.MIMO2:            "mimo2",
    MODEL_ARCH.STEP35:           "step35",
+    MODEL_ARCH.SPARK2_5:         "spark2_5",
    MODEL_ARCH.LLAMA_EMBED:      "llama-embed",
    MODEL_ARCH.MAINCODER:        "maincoder",
    MODEL_ARCH.KIMI_LINEAR:      "kimi-linear",
@@ -2290,6 +2297,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2310,6 +2318,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2333,6 +2342,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2353,6 +2363,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2398,6 +2409,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2500,6 +2512,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_TYPES,
        MODEL_TENSOR.ATTN_NORM_2,
        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -2528,6 +2541,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2557,6 +2571,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2569,6 +2584,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2596,6 +2612,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2627,6 +2644,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2642,6 +2660,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2657,6 +2676,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2671,6 +2691,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2685,6 +2706,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -2705,6 +2727,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -2721,6 +2744,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -2776,6 +2800,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -2792,6 +2817,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -2932,6 +2958,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3065,6 +3092,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3080,6 +3108,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3098,6 +3127,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.ROPE_FACTORS_LONG,
        MODEL_TENSOR.ROPE_FACTORS_SHORT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3135,6 +3165,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3147,6 +3178,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
    MODEL_ARCH.GEMMA2: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3163,6 +3195,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -3181,6 +3214,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -3217,6 +3251,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -3272,6 +3307,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.DENSE_2_OUT,
        MODEL_TENSOR.DENSE_3_OUT,
        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -3292,6 +3328,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3455,6 +3492,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3484,6 +3522,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3498,6 +3537,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3512,6 +3552,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3563,6 +3604,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
    MODEL_ARCH.OLMO: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3575,6 +3617,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3590,6 +3633,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
    MODEL_ARCH.SEED_OSS: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3606,6 +3650,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3656,6 +3701,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3677,6 +3723,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3739,6 +3786,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_A,
        MODEL_TENSOR.ATTN_Q_B,
@@ -3861,6 +3909,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -3937,6 +3986,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4082,6 +4132,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4096,6 +4147,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4117,6 +4169,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.SSM_D,
        MODEL_TENSOR.SSM_NORM,
        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4136,6 +4189,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.SSM_D,
        MODEL_TENSOR.SSM_NORM,
        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4166,6 +4220,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4181,6 +4236,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4206,6 +4262,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4237,6 +4294,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4251,6 +4309,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4276,6 +4335,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.SSM_D,
        MODEL_TENSOR.SSM_NORM,
        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4339,6 +4399,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4378,6 +4439,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4469,6 +4531,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4532,6 +4595,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4547,6 +4611,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4598,6 +4663,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4612,6 +4678,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4629,6 +4696,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.ATTN_NORM,

        # Attention components
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,         # Query projection
        MODEL_TENSOR.ATTN_K,         # Key projection
        MODEL_TENSOR.ATTN_V,         # Value projection
@@ -4661,6 +4729,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4681,6 +4750,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4697,6 +4767,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4739,12 +4810,55 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
    ],
+    MODEL_ARCH.HY_V4: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.HC_HEAD_FN,
+        MODEL_TENSOR.HC_HEAD_BASE,
+        MODEL_TENSOR.HC_HEAD_SCALE,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_SINKS,
+        MODEL_TENSOR.ATTN_Q_A,
+        MODEL_TENSOR.ATTN_Q_A_NORM,
+        MODEL_TENSOR.ATTN_Q_B,
+        MODEL_TENSOR.ATTN_KV_A_MQA,
+        MODEL_TENSOR.ATTN_KV_A_NORM,
+        MODEL_TENSOR.ATTN_K_B,
+        MODEL_TENSOR.ATTN_V_B,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_GATE,
+        MODEL_TENSOR.INDEXER_K_NORM,
+        MODEL_TENSOR.INDEXER_PROJ,
+        MODEL_TENSOR.INDEXER_ATTN_K,
+        MODEL_TENSOR.INDEXER_ATTN_Q_B,
+        MODEL_TENSOR.HC_ATTN_FN,
+        MODEL_TENSOR.HC_ATTN_BASE,
+        MODEL_TENSOR.HC_ATTN_SCALE,
+        MODEL_TENSOR.HC_FFN_FN,
+        MODEL_TENSOR.HC_FFN_BASE,
+        MODEL_TENSOR.HC_FFN_SCALE,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+    ],
    MODEL_ARCH.SMOLLM3: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4761,6 +4875,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4784,6 +4899,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.ATTN_NORM, # operator_norm
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4804,6 +4920,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.ATTN_NORM, # operator_norm
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4819,6 +4936,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4838,6 +4956,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4855,6 +4974,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -4872,6 +4992,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4910,6 +5031,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4973,6 +5095,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -4990,6 +5113,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -5005,6 +5129,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -5158,6 +5283,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -5185,12 +5311,26 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
    ],
+    MODEL_ARCH.SPARK2_5: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_GATE,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
    MODEL_ARCH.LLAMA_EMBED: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -5210,6 +5350,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_Q_NORM,
        MODEL_TENSOR.ATTN_K,
@@ -5226,6 +5367,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -5438,6 +5580,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.ROPE_FREQS,
        MODEL_TENSOR.ATTN_ROT_EMBD,
    ],
+    MODEL_ARCH.HY_V4: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
    MODEL_ARCH.CHATGLM: [
        MODEL_TENSOR.ROPE_FREQS,
    ],
@@ -841,6 +841,9 @@ class GGUFWriter:
        else:
            self.add_array(key, value)

+    def add_recurrent_layers(self, value: Sequence[bool]) -> None:
+        self.add_array(Keys.Attention.RECURRENT_LAYERS.format(arch=self.arch), value)
+
    def add_rope_pattern(self, value: Sequence[bool]) -> None:
        self.add_array(Keys.Attention.ROPE_PATTERN.format(arch=self.arch), value)

@@ -1055,6 +1058,9 @@ class GGUFWriter:
    def add_hyper_connection_epsilon(self, value: float) -> None:
        self.add_float32(Keys.HyperConnection.EPSILON.format(arch=self.arch), value)

+    def add_hyper_connection_magnitude(self, value: float) -> None:
+        self.add_float32(Keys.HyperConnection.MAGNITUDE.format(arch=self.arch), value)
+
    def add_hyper_connection_low_rank(self, value: int) -> None:
        self.add_uint32(Keys.HyperConnection.LOW_RANK.format(arch=self.arch), value)

@@ -385,6 +385,7 @@ class TensorNameMap:
        MODEL_TENSOR.ATTN_SINKS: (
            "model.layers.{bid}.self_attn.sinks", # openai-moe
            "model.layers.{bid}.self_attn.attention_sink_bias", # mimov2
+            "model.layers.{bid}.self_attn.learnable_sink_param", # hy-v4
        ),

        MODEL_TENSOR.ATTN_GATE: (
@@ -392,6 +393,7 @@ class TensorNameMap:
            "model.layers.{bid}.linear_attn.in_proj_z",  # qwen3.5
            "model.layers.{bid}.self_attn.g_proj",    # step3.5 head-wise attention gate
            "model.layers.{bid}.self_attn.output_gate",  # minimax-01
+            "model.layers.{bid}.self_attn.linear_gate",  # hy-v4
        ),

        # Feed-forward norm
@@ -1329,6 +1331,42 @@ class TensorNameMap:
            "model.layers.{bid}.self_attn.index_q_norm", # MSA
        ),

+        MODEL_TENSOR.HC_ATTN_FN: (
+            "model.layers.{bid}.hc_attn_layer.hc_pre.hc_fn", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_ATTN_BASE: (
+            "model.layers.{bid}.hc_attn_layer.hc_pre.hc_base", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_ATTN_SCALE: (
+            "model.layers.{bid}.hc_attn_layer.hc_pre.hc_scale", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_FFN_FN: (
+            "model.layers.{bid}.hc_mlp_layer.hc_pre.hc_fn", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_FFN_BASE: (
+            "model.layers.{bid}.hc_mlp_layer.hc_pre.hc_base", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_FFN_SCALE: (
+            "model.layers.{bid}.hc_mlp_layer.hc_pre.hc_scale", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_HEAD_FN: (
+            "model.hc_head.hc_head_fn",  # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_HEAD_BASE: (
+            "model.hc_head.hc_head_base", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_HEAD_SCALE: (
+            "model.hc_head.hc_head_scale", # hy-v4
+        ),
+
        ############################################################################
        # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
        MODEL_TENSOR.ENC_OUTPUT_NORM: (
@@ -23,4 +23,6 @@ These templates can be updated with the following commands:
 ./scripts/get_chat_template.py Qwen/Qwen3-0.6B                               > models/templates/Qwen-Qwen3-0.6B.jinja
 ./scripts/get_chat_template.py zai-org/GLM-4.5                               > models/templates/zai-org-GLM-4.5.jinja
 ./scripts/get_chat_template.py deepseek-ai/DeepSeek-V3.1                     > models/templates/deepseek-ai-DeepSeek-V3.1.jinja
+./scripts/get_chat_template.py XHToken/Spark-X2.5-1.7B                       > models/templates/Spark2.5.jinja
+./scripts/get_chat_template.py XHToken/Spark-X2.5-4B                         > models/templates/Spark2.5.jinja
 ```
@@ -0,0 +1,110 @@
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+
+{%- set enable_thinking = enable_thinking | default(true) %}
+
+{#- Render a string or a list of text blocks. -#}
+{%- macro render_content(content, context_name) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for block in content %}
+            {%- if block.type == 'text' %}
+                {{- block.text }}
+            {%- else %}
+                {{- raise_exception('Unsupported ' ~ context_name ~ ' content block type: ' ~ (block.type | string)) }}
+            {%- endif %}
+        {%- endfor %}
+    {%- else %}
+        {{- raise_exception(context_name ~ ' content must be a string or a list of text blocks') }}
+    {%- endif %}
+{%- endmacro %}
+
+{#- Default system prompt. -#}
+{%- set default_system = 'you are a helpful assistant.' %}
+
+{#- The first message-level system is placed in the initial system block. -#}
+{%- set ns = namespace(initial_system='') %}
+{%- if messages[0].role == 'system' %}
+    {%- set ns.initial_system = render_content(messages[0].content, 'system') %}
+{%- endif %}
+
+{#- System block. -#}
+{{- '<｜start▁of▁sentence｜><|System|>' + '\n' + default_system }}
+{%- if tools %}
+    {{- '## Tools' + '\n' + 'You have access to the following functions:' + '\n' + '<tools>' }}
+    {%- for tool in tools %}
+        {{- '\n' + tool.function | tojson }}
+    {%- endfor %}
+    {{- '\n' + '</tools>' }}
+{%- endif %}
+{%- if ns.initial_system %}
+    {{- '\n\n' + ns.initial_system }}
+{%- endif %}
+{{- '<｜end▁of▁sentence｜>' }}
+
+{#- Conversation turns. -#}
+{%- for message in messages %}
+    {%- if message.role == 'system' %}
+        {#- The first system message was consumed by the initial block. -#}
+        {%- if not loop.first %}
+            {{- '<｜start▁of▁sentence｜><|System|>\n' + render_content(message.content, 'system') + '<｜end▁of▁sentence｜>' }}
+        {%- endif %}
+    {%- elif message.role == 'user' %}
+        {{- '<｜start▁of▁sentence｜><|User|>' + render_content(message.content, 'user') + '<｜end▁of▁sentence｜>' }}
+    {%- elif message.role == 'assistant' %}
+        {%- set assistant_content = render_content(message.content, 'assistant') %}
+        {%- if message.reasoning_content is defined and message.reasoning_content %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- set reasoning_content = '' %}
+        {%- endif %}
+        {{- '<｜start▁of▁sentence｜><|Bot|>' }}
+        {%- if reasoning_content %}
+            {{- '<think>' + reasoning_content + '</think>' }}
+        {%- else %}
+            {{- '</think>' }}
+        {%- endif %}
+        {%- if assistant_content %}
+            {{- assistant_content }}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is not none %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function.arguments is not mapping %}
+                    {{- raise_exception('tool_call.function.arguments must be a dictionary; normalize JSON strings before apply_chat_template') }}
+                {%- endif %}
+                {%- set args = tool_call.function.arguments %}
+                {{- '<tool_call>' + tool_call.function.name }}
+                {%- for k, v in args.items() %}
+                    {{- '<arg_key>' ~ k ~ '</arg_key><arg_value>' ~ (v if v is string else v | tojson) ~ '</arg_value>' }}
+                {%- endfor %}
+                {{- '</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<｜end▁of▁sentence｜>' }}
+    {%- elif message.role == 'tool' %}
+        {%- if loop.previtem is undefined or loop.previtem.role != 'tool' %}
+            {{- '<｜start▁of▁sentence｜><|Tool|>' }}
+        {%- endif %}
+        {{- '<tool_response>' ~ message.content ~ '</tool_response>' }}
+        {%- if loop.nextitem is undefined or loop.nextitem.role != 'tool' %}
+            {{- '<｜end▁of▁sentence｜>' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unsupported message role: ' ~ message.role) }}
+    {%- endif %}
+{%- endfor %}
+
+{#- Generation prompt. -#}
+{%- if add_generation_prompt %}
+    {{- '<｜start▁of▁sentence｜><|Bot|>' }}
+    {%- if enable_thinking is defined and enable_thinking %}
+        {{- '<think>' }}
+    {%- endif %}
+    {%- if enable_thinking is defined and not enable_thinking %}
+        {{- '</think>' }}
+    {%- endif %}
+{%- endif %}
@@ -120,6 +120,51 @@ else
    fi
 fi

+echo "Checking container images for commit ${SHA}..."
+NIGHTLY_TAG="$(git tag --points-at "${SHA}" | grep -E '(^|-)b[0-9]+(-[0-9a-f]{7})?$' | head -n 1 || true)"
+if [[ -z "${NIGHTLY_TAG}" ]]; then
+    echo "Warning: no nightly tag points at ${SHA} - skipping container image check"
+elif [[ -z "${GITHUB_REPOSITORY:-}" ]]; then
+    echo "Warning: GITHUB_REPOSITORY not set - skipping container image check (local run)"
+else
+    CONTAINER_REPO="${GITHUB_REPOSITORY,,}"  # lower-case owner/repo for ghcr.io
+    GHCR_TOKEN="$(curl -fsSL \
+        "https://ghcr.io/token?scope=repository:${CONTAINER_REPO}:pull&service=ghcr.io" \
+        | grep -oP '"token"\s*:\s*"\K[^"]+')"
+
+    VARIANTS=("" "-cuda" "-cuda13" "-vulkan" "-rocm" "-intel" "-musa" "-openvino")
+    TYPES=("full" "light" "server")
+    CONTAINER_ERR=""
+    for type in "${TYPES[@]}"; do
+        for variant in "${VARIANTS[@]}"; do
+            tag="${type}${variant}-${NIGHTLY_TAG}"
+            STATUS="$(curl -s -o /dev/null -w "%{http_code}" \
+                -H "Authorization: Bearer ${GHCR_TOKEN}" \
+                -H "Accept: application/vnd.oci.image.index.v1+json,application/vnd.docker.distribution.manifest.list.v2+json" \
+                "https://ghcr.io/v2/${CONTAINER_REPO}/manifests/${tag}")"
+            if [[ "${STATUS}" == "200" ]]; then
+                echo "  ${tag} - OK"
+            else
+                echo "  ${tag} - MISSING"
+                CONTAINER_ERR+=" ${tag}"
+            fi
+        done
+    done
+
+    if [[ -n "${CONTAINER_ERR}" ]]; then
+        if [[ "$DRY_RUN" == "true" ]]; then
+            echo "Warning: missing container images for ${NIGHTLY_TAG}:${CONTAINER_ERR} (dry run, continuing)."
+            CHECKS_PASSED=false
+        else
+            echo "Error: missing container images for ${NIGHTLY_TAG}:${CONTAINER_ERR}"
+            echo "The Docker workflow must complete successfully before making a release."
+            exit 1
+        fi
+    else
+        echo "All container images found for ${NIGHTLY_TAG} - OK"
+    fi
+fi
+
 if [[ -n "${GITHUB_OUTPUT:-}" ]]; then
    echo "checks_passed=${CHECKS_PASSED}" >> "$GITHUB_OUTPUT"
 fi
@@ -15,7 +15,6 @@ set(HF_BUCKET         "" CACHE STRING "Hugging Face bucket name")
 set(HF_VERSION        "" CACHE STRING "Version to download (empty = resolve from git)")
 set(HF_ENABLED        "" CACHE STRING "Whether to allow HF Bucket download (ON/OFF)")
 set(BUILD_UI          "" CACHE STRING "Build UI via npm (ON/OFF)")
-set(LLAMA_UI_EMBED    "" CACHE STRING "Path to llama-ui-embed helper")
 set(LLAMA_UI_GZIP     "" CACHE STRING "Apply gzip compress to assets to save bandwidth")

 set(DIST_DIR     "${UI_BINARY_DIR}/dist")
@@ -25,6 +24,223 @@ set(STAMP_FILE   "${UI_BINARY_DIR}/.ui-stamp")
 set(UI_CPP       "${UI_BINARY_DIR}/ui.cpp")
 set(UI_H         "${UI_BINARY_DIR}/ui.h")

+function(mime_from_ext name out_var)
+    string(FIND "${name}" "." ext REVERSE)
+    if(ext GREATER -1)
+        string(SUBSTRING "${name}" ${ext} -1 ext_full)
+        string(SUBSTRING "${ext_full}" 1 -1 ext_str)
+    else()
+        set(ext_str "")
+    endif()
+    if(ext_str STREQUAL "html")
+        set(m "text/html; charset=utf-8")
+    elseif(ext_str STREQUAL "css")
+        set(m "text/css")
+    elseif(ext_str STREQUAL "js")
+        set(m "application/javascript")
+    elseif(ext_str STREQUAL "json")
+        set(m "application/json")
+    elseif(ext_str STREQUAL "webmanifest")
+        set(m "application/manifest+json")
+    elseif(ext_str STREQUAL "svg")
+        set(m "image/svg+xml")
+    elseif(ext_str STREQUAL "png")
+        set(m "image/png")
+    elseif(ext_str STREQUAL "jpg" OR ext_str STREQUAL "jpeg")
+        set(m "image/jpeg")
+    elseif(ext_str STREQUAL "ico")
+        set(m "image/x-icon")
+    elseif(ext_str STREQUAL "woff")
+        set(m "font/woff")
+    elseif(ext_str STREQUAL "woff2")
+        set(m "font/woff2")
+    else()
+        set(m "application/octet-stream")
+    endif()
+    set(${out_var} "${m}" PARENT_SCOPE)
+endfunction()
+
+# Fail when a dist tree is present but is missing files the UI needs at
+# runtime; catches truncated/stale asset trees early with a useful message.
+function(ui_validate_assets files in_dir)
+    list(LENGTH files n_assets)
+    if(n_assets EQUAL 0)
+        return()
+    endif()
+
+    set(found_index FALSE)
+    set(found_manifest FALSE)
+    set(found_sw FALSE)
+    set(found_build_json FALSE)
+    set(found_version_json FALSE)
+    set(found_bundle_js FALSE)
+    set(found_bundle_css FALSE)
+    set(found_workbox_js FALSE)
+
+    foreach(f ${files})
+        get_filename_component(base "${f}" NAME)
+        if(base STREQUAL "index.html")
+            set(found_index TRUE)
+        elseif(base STREQUAL "manifest.webmanifest")
+            set(found_manifest TRUE)
+        elseif(base STREQUAL "sw.js")
+            set(found_sw TRUE)
+        elseif(base STREQUAL "build.json")
+            set(found_build_json TRUE)
+        elseif(base STREQUAL "version.json")
+            set(found_version_json TRUE)
+        elseif(base MATCHES "^bundle.*\\.js$")
+            set(found_bundle_js TRUE)
+        elseif(base MATCHES "^bundle.*\\.css$")
+            set(found_bundle_css TRUE)
+        elseif(base MATCHES "^workbox.*\\.js$")
+            set(found_workbox_js TRUE)
+        endif()
+    endforeach()
+
+    set(missing "")
+    if(NOT found_index)
+        list(APPEND missing "index.html")
+    endif()
+    if(NOT found_manifest)
+        list(APPEND missing "manifest.webmanifest")
+    endif()
+    if(NOT found_sw)
+        list(APPEND missing "sw.js")
+    endif()
+    if(NOT found_build_json)
+        list(APPEND missing "build.json")
+    endif()
+    if(NOT found_version_json)
+        list(APPEND missing "version.json")
+    endif()
+    if(NOT found_bundle_js)
+        list(APPEND missing "bundle[hash].js")
+    endif()
+    if(NOT found_bundle_css)
+        list(APPEND missing "bundle[hash].css")
+    endif()
+    if(NOT found_workbox_js)
+        list(APPEND missing "workbox[hash].js")
+    endif()
+
+    if(missing)
+        set(listing "")
+        foreach(f ${files})
+            string(APPEND listing "    ${f}\n")
+        endforeach()
+        set(missing_list "")
+        foreach(m ${missing})
+            string(APPEND missing_list "    ${m}\n")
+        endforeach()
+        message(FATAL_ERROR
+            "UI: current asset files:\n${listing}"
+            "UI: missing required asset(s):\n${missing_list}"
+            "UI: hint: try cleaning your build directory: ${in_dir}")
+    endif()
+endfunction()
+
+# Generate ui.cpp/ui.h embedding every file of ${dist_dir} (empty table when
+# it has no index.html). When LLAMA_UI_GZIP is enabled, assets are compressed
+# first and served pre-gzipped (llama_ui_use_gzip()).
+function(emit_files dist_dir)
+    set(embed_dir "${dist_dir}")
+    set(use_gzip FALSE)
+
+    if(EXISTS "${dist_dir}/index.html")
+        if(EXISTS "${dist_dir}/_gzip")
+            # a _gzip tree inside dist_dir can only be a leftover from an
+            # older version of this script that staged it there
+            file(REMOVE_RECURSE "${dist_dir}/_gzip")
+            message(STATUS "UI: removed stale gzip tree ${dist_dir}/_gzip")
+        endif()
+        if(LLAMA_UI_GZIP)
+            # Compress every asset into a parallel _gzip/ tree under the build
+            # directory (never write into the source or dist tree); the
+            # structure stays the same: /abc/def --> /_gzip/abc/def.
+            # FORMAT raw produces a bare gzip stream (no archive container)
+            # that can be served with Content-Encoding: gzip. SOURCE_DATE_EPOCH
+            # zeroes the header timestamp so identical inputs give identical
+            # bytes (and therefore stable ETags) on every machine.
+            if(NOT DEFINED ENV{SOURCE_DATE_EPOCH})
+                set(ENV{SOURCE_DATE_EPOCH} 0)
+            endif()
+            set(gzip_root "${UI_BINARY_DIR}/ui-gzip")
+            set(gzip_dir  "${gzip_root}/_gzip")
+            file(REMOVE_RECURSE "${gzip_root}")
+            file(GLOB_RECURSE all_files RELATIVE "${dist_dir}" "${dist_dir}/*")
+            list(FILTER all_files EXCLUDE REGEX "^_gzip/")
+            foreach(f ${all_files})
+                get_filename_component(asset_path "${dist_dir}/${f}" REALPATH)
+                get_filename_component(dst_dir "${gzip_dir}/${f}" DIRECTORY)
+                file(MAKE_DIRECTORY "${dst_dir}")
+                file(ARCHIVE_CREATE
+                    OUTPUT "${gzip_dir}/${f}"
+                    PATHS "${asset_path}"
+                    FORMAT raw
+                    COMPRESSION GZip
+                )
+            endforeach()
+            message(STATUS "UI: gzip compression applied (${gzip_dir})")
+            set(embed_dir "${gzip_dir}")
+            set(use_gzip TRUE)
+        endif()
+    endif()
+
+    set(assets "")
+    if(EXISTS "${embed_dir}/index.html")
+        file(GLOB_RECURSE assets RELATIVE "${embed_dir}" "${embed_dir}/*")
+        list(FILTER assets EXCLUDE REGEX "^_gzip/")
+        list(SORT assets)
+        ui_validate_assets("${assets}" "${embed_dir}")
+    endif()
+
+    list(LENGTH assets n_assets)
+
+    # Only the per-asset data arrays and table rows are built here; all
+    # static C++ lives in the ui.h.in / ui.cpp.in templates. configure_file
+    # rewrites an output only when its contents change, so the library is
+    # not recompiled needlessly. @ONLY keeps ${...} in the content literal;
+    # mime types come from a fixed list.
+    set(ASSET_ARRAYS "")
+    set(ASSET_TABLE "")
+    set(idx 0)
+
+    foreach(f IN LISTS assets)
+        file(READ "${embed_dir}/${f}" hex HEX)
+        if(hex STREQUAL "")
+            message(FATAL_ERROR "UI: empty file: ${embed_dir}/${f}")
+        endif()
+
+        string(REGEX REPLACE "(..)" "0x\\1," bytes "${hex}")
+        file(SHA256 "${embed_dir}/${f}" etag)
+        mime_from_ext("${f}" mime)
+
+        string(APPEND ASSET_ARRAYS
+            "static const unsigned char asset_${idx}[] = {${bytes}};\n")
+
+        string(APPEND ASSET_TABLE
+            "    { \"${f}\", asset_${idx}, sizeof(asset_${idx}), \"\\\"${etag}\\\"\", \"${mime}\" },\n")
+
+        math(EXPR idx "${idx} + 1")
+    endforeach()
+
+    set(LLAMA_UI_HAS_ASSETS 0)
+    if(n_assets GREATER 0)
+        set(LLAMA_UI_HAS_ASSETS 1)
+    endif()
+    set(N_ASSETS "${n_assets}")
+    set(USE_GZIP false)
+    if(use_gzip)
+        set(USE_GZIP true)
+    endif()
+
+    set(UI_TEMPLATE_DIR "${LLAMA_SOURCE_DIR}/tools/ui")
+    configure_file("${UI_TEMPLATE_DIR}/ui.h.in"   "${UI_H}"   @ONLY)
+    configure_file("${UI_TEMPLATE_DIR}/ui.cpp.in" "${UI_CPP}" @ONLY)
+    message(STATUS "UI: embedded ${n_assets} assets")
+endfunction()
+
 function(npm_build_should_skip out_var)
    set(${out_var} FALSE PARENT_SCOPE)

@@ -250,48 +466,6 @@ function(hf_download version out_var out_resolved)
    endforeach()
 endfunction()

-function(emit_files dist_dir)
-    # If gzip is requested, compress every asset into a parallel _gzip/ tree
-    # the structure stays the same; for ex: /abc/def --> /_gzip/abc/def
-    # embed.cpp will check for _gzip and will pick it up
-    if(LLAMA_UI_GZIP AND EXISTS "${dist_dir}/index.html")
-        find_program(GZIP_EXECUTABLE gzip)
-        if(NOT GZIP_EXECUTABLE)
-            message(WARNING "UI: LLAMA_UI_GZIP requested but gzip not found, embedding uncompressed")
-        else()
-            set(gzip_dir "${dist_dir}/_gzip")
-            file(REMOVE_RECURSE "${gzip_dir}")
-            file(GLOB_RECURSE all_files RELATIVE "${dist_dir}" "${dist_dir}/*")
-            foreach(f ${all_files})
-                get_filename_component(dst_dir "${gzip_dir}/${f}" DIRECTORY)
-                file(MAKE_DIRECTORY "${dst_dir}")
-                execute_process(
-                    COMMAND "${GZIP_EXECUTABLE}" -c "${dist_dir}/${f}"
-                    OUTPUT_FILE "${gzip_dir}/${f}"
-                    RESULT_VARIABLE gz_rc
-                )
-                if(NOT gz_rc EQUAL 0)
-                    message(FATAL_ERROR "UI: gzip failed for ${f}")
-                endif()
-            endforeach()
-            message(STATUS "UI: gzip compression applied (${gzip_dir})")
-        endif()
-    endif()
-
-    set(args "${UI_CPP}" "${UI_H}")
-    if(EXISTS "${dist_dir}/index.html")
-        list(APPEND args "${dist_dir}")
-    endif()
-
-    execute_process(
-        COMMAND "${LLAMA_UI_EMBED}" ${args}
-        RESULT_VARIABLE rc
-    )
-    if(NOT rc EQUAL 0)
-        message(FATAL_ERROR "UI: llama-ui-embed failed (${rc})")
-    endif()
-endfunction()
-
 # ---------------------------------------------------------------------------
 # 1. Priority 1: pre-built assets supplied in tools/ui/dist
 # ---------------------------------------------------------------------------
@@ -121,6 +121,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_HUNYUAN_DENSE,    "hunyuan-dense"    },
    { LLM_ARCH_HUNYUAN_VL,       "hunyuan_vl"       },
    { LLM_ARCH_HY_V3,            "hy_v3"            },
+    { LLM_ARCH_HY_V4,            "hy_v4"            },
    { LLM_ARCH_SMOLLM3,          "smollm3"          },
    { LLM_ARCH_OPENAI_MOE,       "gpt-oss"          },
    { LLM_ARCH_LFM2,             "lfm2"             },
@@ -145,6 +146,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_PADDLEOCR,        "paddleocr"        },
    { LLM_ARCH_MIMO2,            "mimo2"            },
    { LLM_ARCH_STEP35,           "step35"           },
+    { LLM_ARCH_SPARK2_5,         "spark2_5"         },
    { LLM_ARCH_LLAMA_EMBED,      "llama-embed"      },
    { LLM_ARCH_MAINCODER,        "maincoder"        },
    { LLM_ARCH_KIMI_LINEAR,      "kimi-linear"      },
@@ -294,6 +296,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_HYPER_CONNECTION_COUNT,                 "%s.hyper_connection.count"                 },
    { LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS,   "%s.hyper_connection.sinkhorn_iterations"   },
    { LLM_KV_HYPER_CONNECTION_EPSILON,               "%s.hyper_connection.epsilon"               },
+    { LLM_KV_HYPER_CONNECTION_MAGNITUDE,             "%s.hyper_connection.magnitude"             },
    { LLM_KV_HYPER_CONNECTION_LOW_RANK,              "%s.hyper_connection.low_rank"              },

    { LLM_KV_PLE_LAYERS,                             "%s.ple.layers"                             },
@@ -1130,6 +1133,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
        case LLM_ARCH_OLMOE:
        case LLM_ARCH_DEEPSEEK2:
        case LLM_ARCH_DEEPSEEK32:
+        case LLM_ARCH_HY_V4:
        case LLM_ARCH_DOTS3NOTE:
        case LLM_ARCH_GLM_DSA:
        case LLM_ARCH_BITNET:
@@ -126,6 +126,7 @@ enum llm_arch {
    LLM_ARCH_HUNYUAN_DENSE,
    LLM_ARCH_HUNYUAN_VL,
    LLM_ARCH_HY_V3,
+    LLM_ARCH_HY_V4,
    LLM_ARCH_SMOLLM3,
    LLM_ARCH_OPENAI_MOE,
    LLM_ARCH_LFM2,
@@ -146,6 +147,7 @@ enum llm_arch {
    LLM_ARCH_PADDLEOCR,
    LLM_ARCH_MIMO2,
    LLM_ARCH_STEP35,
+    LLM_ARCH_SPARK2_5,
    LLM_ARCH_LLAMA_EMBED,
    LLM_ARCH_MAINCODER,
    LLM_ARCH_KIMI_LINEAR,
@@ -299,6 +301,7 @@ enum llm_kv {
    LLM_KV_HYPER_CONNECTION_COUNT,
    LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS,
    LLM_KV_HYPER_CONNECTION_EPSILON,
+    LLM_KV_HYPER_CONNECTION_MAGNITUDE,
    LLM_KV_HYPER_CONNECTION_LOW_RANK,

    LLM_KV_PLE_LAYERS,
@@ -2317,7 +2317,8 @@ uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
        (model.arch == LLM_ARCH_DFLASH && model.hparams.dsv4_hc_mult > 0) ||
        model.arch == LLM_ARCH_NANBEIGE ||
        model.arch == LLM_ARCH_MINIMAX_01 ||
-        model.arch == LLM_ARCH_MINIMAX_M3) {
+        model.arch == LLM_ARCH_MINIMAX_M3 ||
+        model.arch == LLM_ARCH_HY_V4) {
        res = std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
    } else if (model.arch == LLM_ARCH_DFLASH && model.hparams.dflash_selector_rank > 0) {
        // DFlash2's convolutions and selector are shape work rather than matmuls,
@@ -492,7 +492,7 @@ const char * llama_grammar_parser::parse_sequence(
            total_rules = min_times;
        }

-        if (n_prev_rules * total_rules >= MAX_REPETITION_THRESHOLD) {
+        if (n_prev_rules * total_rules > MAX_REPETITION_THRESHOLD) {
            throw std::runtime_error("number of rules that are going to be repeated multiplied by the new repetition exceeds sane defaults, please reduce the number of repetitions or rule complexity");
        }

@@ -566,7 +566,10 @@ void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {

    mctx->get_lid()->set_input_kq_mask(self_kq_mask_lid, ubatch, cparams.causal_attn);

-    mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
+    // left unallocated when the indexer does not use the rotation
+    if (self_k_rot_lid && self_k_rot_lid->buffer) {
+        mctx->get_lid()->set_input_k_rot(self_k_rot_lid);
+    }
 }

 bool llm_graph_input_attn_k_dsa::can_reuse(const llm_graph_params & params) {
@@ -1620,8 +1623,26 @@ llm_graph_qkv llm_graph_context::build_qkv(
                  int64_t   n_head,
                  int64_t   n_head_kv,
                      int   il) const {
-    const int64_t n_embd_q  = n_embd_head * n_head;
-    const int64_t n_embd_kv = n_embd_head * n_head_kv;
+    return build_qkv(layer, cur,
+            n_embd_head, n_head,
+            n_embd_head, n_head_kv,
+            n_embd_head, n_head_kv,
+            il);
+}
+
+llm_graph_qkv llm_graph_context::build_qkv(
+        const llama_layer & layer,
+              ggml_tensor * cur,
+                  int64_t   n_embd_head_q,
+                  int64_t   n_head_q,
+                  int64_t   n_embd_head_k,
+                  int64_t   n_head_k,
+                  int64_t   n_embd_head_v,
+                  int64_t   n_head_v,
+                      int   il,
+                     bool   reshape) const {
+    const int64_t n_embd_q = n_embd_head_q * n_head_q;
+    const int64_t n_embd_k = n_embd_head_k * n_head_k;

    ggml_tensor * Qcur, * Kcur, * Vcur;

@@ -1632,59 +1653,93 @@ llm_graph_qkv llm_graph_context::build_qkv(
        if (layer.wqkv_b) {
            qkv = ggml_add(ctx0, qkv, layer.wqkv_b);
            cb(qkv, "wqkv_b", il);
+        } else if (layer.wq_b && layer.wk_b && layer.wv_b) {
+            // Fused weights may coexist with separate Q/K/V biases in legacy or custom GGUFs.
+            ggml_tensor * qkv_b = ggml_concat(ctx0, ggml_concat(ctx0, layer.wq_b, layer.wk_b, 0), layer.wv_b, 0);
+            qkv = ggml_add(ctx0, qkv, qkv_b);
+            cb(qkv, "wqkv_b", il);
        }
-        if (hparams.f_clamp_kqv > 0.0f) {
+        if (reshape && hparams.f_clamp_kqv > 0.0f) {
            qkv = ggml_clamp(ctx0, qkv, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
            cb(qkv, "wqkv_clamped", il);
        }
-        Qcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head,    n_tokens,
-            ggml_row_size(qkv->type, n_embd_head), qkv->nb[1], 0);
-        Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens,
-            ggml_row_size(qkv->type, n_embd_head), qkv->nb[1],
-            ggml_row_size(qkv->type, n_embd_q));
-        Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens,
-            ggml_row_size(qkv->type, n_embd_head), qkv->nb[1],
-            ggml_row_size(qkv->type, n_embd_q + n_embd_kv));
+        if (reshape) {
+            Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head_q, n_tokens,
+                ggml_row_size(qkv->type, n_embd_head_q), qkv->nb[1], 0);
+            Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_k, n_tokens,
+                ggml_row_size(qkv->type, n_embd_head_k), qkv->nb[1],
+                ggml_row_size(qkv->type, n_embd_q));
+            Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_v, n_tokens,
+                ggml_row_size(qkv->type, n_embd_head_v), qkv->nb[1],
+                ggml_row_size(qkv->type, n_embd_q + n_embd_k));
+        } else {
+            Qcur = ggml_view_2d(ctx0, qkv, n_embd_q, n_tokens, qkv->nb[1], 0);
+            Kcur = ggml_view_2d(ctx0, qkv, n_embd_k, n_tokens, qkv->nb[1],
+                ggml_row_size(qkv->type, n_embd_q));
+            Vcur = ggml_view_2d(ctx0, qkv, n_embd_head_v * n_head_v, n_tokens, qkv->nb[1],
+                ggml_row_size(qkv->type, n_embd_q + n_embd_k));
+        }
+        if (!reshape) {
+            Qcur = ggml_cont(ctx0, Qcur);
+            Kcur = ggml_cont(ctx0, Kcur);
+            Vcur = ggml_cont(ctx0, Vcur);
+        }
    } else {
        // separate Q/K/V path
        Qcur = build_lora_mm(layer.wq, cur, layer.wq_s);
-        cb(Qcur, "Qcur", il);
-        if (layer.wq_b) {
-            Qcur = ggml_add(ctx0, Qcur, layer.wq_b);
+        if (reshape) {
            cb(Qcur, "Qcur", il);
        }
-        if (hparams.f_clamp_kqv > 0.0f) {
+        if (layer.wq_b) {
+            Qcur = ggml_add(ctx0, Qcur, layer.wq_b);
+            if (reshape) {
+                cb(Qcur, "Qcur", il);
+            }
+        }
+        if (reshape && hparams.f_clamp_kqv > 0.0f) {
            Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
            cb(Qcur, "Qcur_clamped", il);
        }
        Kcur = build_lora_mm(layer.wk, cur, layer.wk_s);
-        cb(Kcur, "Kcur", il);
-        if (layer.wk_b) {
-            Kcur = ggml_add(ctx0, Kcur, layer.wk_b);
+        if (reshape) {
            cb(Kcur, "Kcur", il);
        }
-        if (hparams.f_clamp_kqv > 0.0f) {
+        if (layer.wk_b) {
+            Kcur = ggml_add(ctx0, Kcur, layer.wk_b);
+            if (reshape) {
+                cb(Kcur, "Kcur", il);
+            }
+        }
+        if (reshape && hparams.f_clamp_kqv > 0.0f) {
            Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
            cb(Kcur, "Kcur_clamped", il);
        }
        Vcur = build_lora_mm(layer.wv, cur, layer.wv_s);
-        cb(Vcur, "Vcur", il);
-        if (layer.wv_b) {
-            Vcur = ggml_add(ctx0, Vcur, layer.wv_b);
+        if (reshape) {
            cb(Vcur, "Vcur", il);
        }
-        if (hparams.f_clamp_kqv > 0.0f) {
+        if (layer.wv_b) {
+            Vcur = ggml_add(ctx0, Vcur, layer.wv_b);
+            if (reshape) {
+                cb(Vcur, "Vcur", il);
+            }
+        }
+        if (reshape && hparams.f_clamp_kqv > 0.0f) {
            Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
            cb(Vcur, "Vcur_clamped", il);
        }
-        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-        Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+        if (reshape) {
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_q, n_head_q, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_k, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_v, n_tokens);
+        }
    }

-    cb(Qcur, "Qcur", il);
-    cb(Kcur, "Kcur", il);
-    cb(Vcur, "Vcur", il);
+    if (reshape) {
+        cb(Qcur, "Qcur", il);
+        cb(Kcur, "Kcur", il);
+        cb(Vcur, "Vcur", il);
+    }

    return { Qcur, Kcur, Vcur };
 }
@@ -2170,7 +2225,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                    const float limit = hparams.swiglu_clamp_exp[il];
                    constexpr float eps = 1e-6f;
                    if (limit > eps) {
-                        if (arch == LLM_ARCH_DEEPSEEK4 || (arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0)) {
+                        if (arch == LLM_ARCH_DEEPSEEK4 || (arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0) || arch == LLM_ARCH_HY_V4) {
                            cur = ggml_swiglu_clamp(ctx0, cur, up, limit);
                        } else {
                            up = ggml_clamp(ctx0, up, -limit, limit);
@@ -1079,6 +1079,19 @@ struct llm_graph_context {
                  int64_t   n_head_kv,
                      int   il) const;

+    // Set reshape to false to return contiguous projections before clamp/reshape.
+    llm_graph_qkv build_qkv(
+        const llama_layer & layer,
+              ggml_tensor * cur,
+                  int64_t   n_embd_head_q,
+                  int64_t   n_head_q,
+                  int64_t   n_embd_head_k,
+                  int64_t   n_head_k,
+                  int64_t   n_embd_head_v,
+                  int64_t   n_head_v,
+                      int   il,
+                     bool   reshape = true) const;
+
    ggml_tensor * build_ffn(
             ggml_tensor * cur,
             ggml_tensor * up,
@@ -297,6 +297,9 @@ struct llama_hparams {
    // 0 = full rank (DeepSeek-V4)
    uint32_t hc_low_rank = 0;

+    // scale of the hyper-connection post gate (DeepSeek-V4 hardcodes 2.0)
+    float    hc_magnitude = 0.0f;
+
    uint32_t ple_ngram_size      = 0;
    uint32_t ple_heads_per_ngram = 0;
    uint32_t ple_conv_kernel     = 0;
@@ -27,6 +27,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
        case LLM_ARCH_APERTUS:
        case LLM_ARCH_MIMO2:
        case LLM_ARCH_STEP35:
+        case LLM_ARCH_SPARK2_5:
        case LLM_ARCH_MUSE_GLIMMER:
        case LLM_ARCH_MELLUM:
        case LLM_ARCH_LAGUNA:
@@ -314,6 +315,7 @@ void llama_model_saver::add_kv_from_model() {
    add_kv(LLM_KV_HYPER_CONNECTION_COUNT,               hparams.dsv4_hc_mult);
    add_kv(LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, hparams.dsv4_hc_sinkhorn_iters);
    add_kv(LLM_KV_HYPER_CONNECTION_EPSILON,             hparams.dsv4_hc_eps);
+    add_kv(LLM_KV_HYPER_CONNECTION_MAGNITUDE,           hparams.hc_magnitude);
    add_kv(LLM_KV_HASH_LAYER_COUNT,                     hparams.dsv4_hash_layer_count);
    add_kv(LLM_KV_HYPER_CONNECTION_LOW_RANK,             hparams.hc_low_rank);

@@ -288,6 +288,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_hunyuan_dense(params);
        case LLM_ARCH_HY_V3:
            return new llama_model_hy_v3(params);
+        case LLM_ARCH_HY_V4:
+            return new llama_model_hy_v4(params);
        case LLM_ARCH_SMOLLM3:
            return new llama_model_smollm3(params);
        case LLM_ARCH_OPENAI_MOE:
@@ -336,6 +338,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_kimi_k3(params);
        case LLM_ARCH_STEP35:
            return new llama_model_step35(params);
+        case LLM_ARCH_SPARK2_5:
+            return new llama_model_spark2_5(params);
        default:
            throw std::runtime_error(std::string("unsupported model architecture: '") + llm_arch_name(arch) + "'");
    }
@@ -2053,7 +2057,8 @@ void llama_model::print_info() const {

        if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_DEEPSEEK2OCR ||
                arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA ||
-                arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4) {
+                arch == LLM_ARCH_DOTS3NOTE || arch == LLM_ARCH_MISTRAL4 ||
+                arch == LLM_ARCH_HY_V4) {
            LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
            LLAMA_LOG_INFO("%s: n_lora_q              = %d\n",     __func__, hparams.n_lora_q);
            LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
@@ -2322,6 +2327,48 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                            nullptr);
                }
            } break;
+        case LLM_ARCH_HY_V4:
+            {
+                if (hparams.indexer_top_k == 0) {
+                    // full-attention checkpoint: no indexer, so no indexer key cache
+                    res = new llama_kv_cache(
+                            *this,
+                            hparams,
+                            params.type_k,
+                            params.type_v,
+                            !cparams.flash_attn,
+                            cparams.offload_kqv,
+                            cparams.kv_unified,
+                            cparams.n_ctx_seq,
+                            cparams.n_seq_max,
+                            1,
+                            hparams.n_swa,
+                            hparams.swa_type,
+                            nullptr,
+                            nullptr,
+                            nullptr,
+                            nullptr);
+                } else {
+                    // only "full" layers own an indexer, so the shared layers need no indexer cache
+                    llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return hparams.is_indexer_full(il); };
+
+                    res = new llama_kv_cache_dsa(
+                            *this,
+                            params.type_k,
+                            params.type_v,
+                            !cparams.flash_attn,
+                            cparams.offload_kqv,
+                            cparams.kv_unified,
+                            cparams.n_ctx_seq,
+                            cparams.n_seq_max,
+                            1,
+                            hparams.n_swa,
+                            hparams.swa_type,
+                            nullptr,
+                            filter_lid,
+                            nullptr);
+                }
+            } break;
        case LLM_ARCH_DOTS3NOTE:
            {
                GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE);
@@ -2881,6 +2928,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_DOTS3NOTE:
        case LLM_ARCH_NANBEIGE:
        case LLM_ARCH_POCKETTTS:
+        // HY_V4 rotates consecutive pairs, matching the reference implementation
+        case LLM_ARCH_HY_V4:
            return LLAMA_ROPE_TYPE_NORM;

        // the pairs of head values are offset by n_rot/2
@@ -2952,6 +3001,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_QWEN3NEXT:
        case LLM_ARCH_MIMO2:
        case LLM_ARCH_STEP35:
+        case LLM_ARCH_SPARK2_5:
        case LLM_ARCH_TALKIE:
        case LLM_ARCH_MELLUM:
            return LLAMA_ROPE_TYPE_NEOX;
@@ -3186,6 +3236,12 @@ void llama_model_base::create_tensor_qkv(llama_layer & layer, int bid,
    layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", bid), {n_embd_, n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
    if (layer.wqkv) {
        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", bid), {n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
+        // Fused weights may coexist with separate Q/K/V biases in legacy or custom GGUFs.
+        if (!layer.wqkv_b) {
+            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", bid), {n_embd_q_}, TENSOR_NOT_REQUIRED);
+            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", bid), {n_embd_k_}, TENSOR_NOT_REQUIRED);
+            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", bid), {n_embd_v_}, TENSOR_NOT_REQUIRED);
+        }
    } else {
        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", bid), {n_embd_, n_embd_q_}, flags);
        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", bid), {n_embd_, n_embd_k_}, flags);
@@ -318,12 +318,21 @@ struct llm_tokenizer_bpe : llm_tokenizer {
            case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM:
            case LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE:
            case LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM:
+            case LLAMA_VOCAB_PRE_TYPE_HY_V4:
                regex_exprs = {
                    "\\p{N}{1,3}",
                    "[一-龥぀-ゟ゠-ヿ]+",
                    "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\r\n]*|\\s*[\r\n]+|\\s+(?!\\S)|\\s+",
                };
                break;
+            case LLAMA_VOCAB_PRE_TYPE_SPARK2_5:
+                regex_exprs = {
+                    "\\p{N}{1,3}",
+                    "[一-龥぀-ゟ゠-ヿ]+",
+                    "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\r\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+|[\r\n]|\\s+(?!\\S)|\\s+",
+                    "\\p{N}",
+                };
+                break;
            case LLAMA_VOCAB_PRE_TYPE_YOUTU:
                regex_exprs = {
                    "[가-힣ㄱ-ㆎ]+|[！…“”‘’—：；，、-〿︰-﹏]+|[ㄅ-ㄯ]+|[一-龥぀-ゟ゠-ヿ]+",
@@ -2169,6 +2178,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                    tokenizer_pre == "deepseek-v3") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM;
                clean_spaces = false;
+            } else if (
+                    tokenizer_pre == "spark2_5") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SPARK2_5;
+                clean_spaces = false;
            } else if (
                    tokenizer_pre == "youtu") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_YOUTU;
@@ -2350,6 +2363,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                tokenizer_pre == "hunyuan-dense") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE;
                clean_spaces = false;
+            } else if (
+                tokenizer_pre == "hy_v4") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_HY_V4;
+                clean_spaces = false;
            } else if (
                tokenizer_pre == "joyai-llm") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM;
@@ -65,6 +65,8 @@ enum llama_vocab_pre_type {
    LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI = 54,
    LLAMA_VOCAB_PRE_TYPE_MELLUM2           = 55,
    LLAMA_VOCAB_PRE_TYPE_LAGUNA            = 56,
+    LLAMA_VOCAB_PRE_TYPE_HY_V4             = 57,
+    LLAMA_VOCAB_PRE_TYPE_SPARK2_5          = 58,
 };

 struct LLM_KV;
@@ -280,8 +280,8 @@ llama_model_bailingmoe3::graph::graph(const llama_model & model, const llm_graph
            ggml_tensor * beta = ggml_mul_mat(ctx0, layer.ssm_beta, cur);
            beta = ggml_sigmoid(ctx0, ggml_reshape_4d(ctx0, beta, 1, n_head, n_seq_tokens, n_seqs));

-            q = ggml_l2_norm(ctx0, q, hparams.f_norm_rms_eps);
-            k = ggml_l2_norm(ctx0, k, hparams.f_norm_rms_eps);
+            q = build_gdn_l2_norm(ctx0, q, hparams.f_norm_rms_eps);
+            k = build_gdn_l2_norm(ctx0, k, hparams.f_norm_rms_eps);

            ggml_tensor * states_all = mctx_cur->get_s_l(il);
            ggml_tensor * state = build_rs(inp_rs, states_all, hparams.n_embd_s(), n_seqs);
@@ -475,21 +475,12 @@ llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_p
            const int ocr_rope_type = GGML_ROPE_TYPE_NEOX;
            GGML_ASSERT(n_embed_head == n_embd_head_k && n_embed_head == n_embd_head_v);

-            ggml_tensor * Qcur = NULL;
-            ggml_tensor * Kcur = NULL;
-            ggml_tensor * Vcur = NULL;
-
-            Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-            Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-            Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embed_head, n_head, n_head, il);
            cb(Qcur, "q", il);
            cb(Kcur, "k", il);
            cb(Vcur, "v", il);

-            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embed_head, n_head, n_tokens);
-            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embed_head, n_head, n_tokens);
-            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embed_head, n_head, n_tokens);
-
            GGML_ASSERT(fabs(freq_base - 10000.0) < 1e-4);
            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_embed_head, ocr_rope_type, 0, freq_base, 1, 0, 1, 0, 0);
            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_embed_head, ocr_rope_type, 0, freq_base, 1, 0, 1, 0, 0);
@@ -40,9 +40,7 @@ void llama_model_deepseek2ocr::load_arch_tensors(llama_model_loader &) {
    for (int i = 0; i < n_layer; ++i) {
        auto & layer = layers[i];

-        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
-        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd}, 0);
-        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd}, 0);
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd, n_embd, 0);
        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);

        // norm
@@ -176,7 +176,14 @@ llama_model_gemma3n::graph::graph(const llama_model & model, const llm_graph_par
                    hparams.f_attention_scale, il);
        } else {
            // reuse KV cache of earlier layers
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            ggml_tensor * Qcur;
+            if (model.layers[il].wqkv) {
+                ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
+                const int64_t q_dim = n_embd_head * n_head;
+                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, q_dim, n_tokens, qkv->nb[1], 0));
+            } else {
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+            }
            cb(Qcur, "Qcur", il);
            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);

@@ -75,9 +75,13 @@ void llama_model_gemma4::load_arch_tensors(llama_model_loader &) {
        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);

        // note: use_alternative_attention (v_proj is optional, if it's not present, use k_proj)
-        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head * n_head}, 0);
-        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k}, kv_flags);
-        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v}, TENSOR_NOT_REQUIRED);
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i),
+            {n_embd, n_embd_head * n_head + n_embd_k + n_embd_v}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
+        if (!layer.wqkv) {
+            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head * n_head}, 0);
+            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k}, kv_flags);
+            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v}, TENSOR_NOT_REQUIRED);
+        }
        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head * n_head, n_embd}, 0);

        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head}, 0);
@@ -202,9 +206,17 @@ llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_para

        // Q projection (shared for both non-KV and KV layers)
        // this is to mirror Gemma4Attention in pytorch code
+        ggml_tensor * qkv_fused = nullptr;
        ggml_tensor * Qcur;
-        {
+        if (model.layers[il].wqkv) {
+            qkv_fused = build_lora_mm(model.layers[il].wqkv, cur, model.layers[il].wqkv_s);
+            cb(qkv_fused, "wqkv", il);
+            const int64_t q_dim = n_embd_head * n_head;
+            Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv_fused, q_dim, n_tokens, qkv_fused->nb[1], 0));
+        } else {
            Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
+        }
+        {
            cb(Qcur, "Qcur", il);

            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
@@ -219,12 +231,22 @@ llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_para

        // self-attention
        if (hparams.has_kv(il)) {
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+            if (qkv_fused) {
+                const int64_t q_dim = n_embd_head * n_head;
+                const int64_t k_dim = n_embd_head * n_head_kv;
+                const int64_t v_dim = n_embd_head * n_head_kv;
+                const size_t  esize = ggml_element_size(qkv_fused);
+                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv_fused, k_dim, n_tokens, qkv_fused->nb[1], q_dim * esize));
+                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv_fused, v_dim, n_tokens, qkv_fused->nb[1], (q_dim + k_dim) * esize));
+            } else {
+                Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
+                Vcur = model.layers[il].wv
+                       ? build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s)
+                       : Kcur; // if v_proj is not present, use Kcur as Vcur
+            }
            cb(Kcur, "Kcur", il);
-
-            ggml_tensor * Vcur = model.layers[il].wv
-                                    ? build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s)
-                                    : Kcur; // if v_proj is not present, use Kcur as Vcur
            cb(Vcur, "Vcur", il);

            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
@@ -0,0 +1,601 @@
+#include "models.h"
+
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-dsa.h"
+
+#include <cmath>
+
+// iHC (independent Hyper-Connections) helpers. Same layout as the DeepSeek-V4 HC, but without
+// the comb/sinkhorn term: hc_fn makes only 2*hc coefficients (pre + post). The streams mix
+// through the pre-reduce / post-distribute round trip instead.
+
+static size_t hy_v4_elem_offset(const ggml_tensor * t, int64_t i) {
+    return ggml_row_size(t->type, i);
+}
+
+static ggml_tensor * hy_v4_view_1d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t i0) {
+    return ggml_view_1d(ctx, t, ne0, hy_v4_elem_offset(t, i0));
+}
+
+static ggml_tensor * hy_v4_view_2d(ggml_context * ctx, ggml_tensor * t, int64_t ne0, int64_t ne1, int64_t i0) {
+    return ggml_view_2d(ctx, t, ne0, ne1, t->nb[1], hy_v4_elem_offset(t, i0));
+}
+
+void llama_model_hy_v4::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,       hparams.n_lora_q);
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,    hparams.n_embd_head_k_mla_impl);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA,  hparams.n_embd_head_v_mla_impl);
+    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm,  false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func,   false);
+
+    // routed-expert SwiGLU logits clamp (shared/dense experts are NOT clamped, so
+    // swiglu_clamp_shexp is intentionally left at its 0 default)
+    ml.get_key_or_arr(LLM_KV_SWIGLU_CLAMP_EXP, hparams.swiglu_clamp_exp, hparams.n_layer_all, false);
+
+    ml.get_key(LLM_KV_HYPER_CONNECTION_COUNT,     hparams.dsv4_hc_mult);
+    ml.get_key(LLM_KV_HYPER_CONNECTION_EPSILON,   hparams.dsv4_hc_eps);
+    ml.get_key(LLM_KV_HYPER_CONNECTION_MAGNITUDE, hparams.hc_magnitude);
+
+    // DSA is absent on the all-full_attention checkpoints, so indexer_top_k stays 0 there
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head,    false);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size, false);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K,      hparams.indexer_top_k,     false);
+
+    if (hparams.indexer_top_k > 0) {
+        // the reference plumbs rms_norm_eps into the indexer k_norm LayerNorm, and build_norm
+        // reads f_norm_eps for LLM_NORM
+        hparams.f_norm_eps = hparams.f_norm_rms_eps;
+
+        if (hparams.indexer_n_head == 0 || hparams.indexer_head_size <= hparams.n_rot()) {
+            throw std::runtime_error("hy_v4: bad indexer head count / key length");
+        }
+
+        ml.get_key_or_arr(LLM_KV_ATTENTION_INDEXER_TYPES, hparams.is_indexer_full_impl, hparams.n_layer(), false);
+        if (!hparams.is_indexer_full(0)) {
+            throw std::runtime_error("hy_v4: layer 0 must own an indexer, nothing precedes it to share");
+        }
+    }
+
+    GGML_ASSERT(hparams.is_mla());
+
+    type = LLM_TYPE_UNKNOWN;
+}
+
+void llama_model_hy_v4::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t n_embd_head_k_mla   = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_v_mla   = hparams.n_embd_head_v_mla();
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+    GGML_ASSERT(n_embd_head_qk_nope >= 1);
+
+    const int64_t q_lora_rank     = hparams.n_lora_q;
+    const int64_t kv_lora_rank    = hparams.n_lora_kv;
+    const int64_t n_ff_exp        = hparams.n_ff_exp();
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+    const int64_t hc              = hparams.dsv4_hc_mult;
+
+    tok_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    // global iHC head (collapses hc streams before the final norm)
+    hc_head_fn    = create_tensor(tn(LLM_TENSOR_HC_HEAD_FN,    "weight"), {hc * n_embd, hc}, 0);
+    hc_head_base  = create_tensor(tn(LLM_TENSOR_HC_HEAD_BASE,  "weight"), {hc}, 0);
+    hc_head_scale = create_tensor(tn(LLM_TENSOR_HC_HEAD_SCALE, "weight"), {1}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm     = create_tensor(tn(LLM_TENSOR_ATTN_NORM,     "weight", i), {n_embd}, 0);
+        layer.attn_sinks    = create_tensor(tn(LLM_TENSOR_ATTN_SINKS,    "weight", i), {n_head}, 0);
+
+        layer.wq_a          = create_tensor(tn(LLM_TENSOR_ATTN_Q_A,      "weight", i), {n_embd, q_lora_rank}, 0);
+        layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
+        layer.wq_b          = create_tensor(tn(LLM_TENSOR_ATTN_Q_B,      "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
+        layer.wkv_a_mqa     = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+        layer.attn_kv_a_norm= create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM,"weight", i), {kv_lora_rank}, 0);
+        layer.wk_b          = create_tensor(tn(LLM_TENSOR_ATTN_K_B,      "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
+        layer.wv_b          = create_tensor(tn(LLM_TENSOR_ATTN_V_B,      "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+        layer.wo            = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
+        layer.wqkv_gate     = create_tensor(tn(LLM_TENSOR_ATTN_GATE,     "weight", i), {n_embd, n_head * n_embd_head_v_mla}, 0);
+
+        // only "full" indexer layers ship weights; "shared" layers reuse their top-k
+        if (hparams.indexer_top_k > 0 && hparams.is_indexer_full(i)) {
+            const int64_t n_indexer_head = hparams.indexer_n_head;
+            const int64_t n_embd_indexer = hparams.indexer_head_size;
+
+            layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, n_indexer_head * n_embd_indexer}, 0);
+            layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, n_embd_indexer}, 0);
+            layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {n_embd_indexer}, 0);
+            layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {n_embd_indexer}, 0);
+            layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, n_indexer_head}, 0);
+        }
+
+        layer.hc_attn_fn    = create_tensor(tn(LLM_TENSOR_HC_ATTN_FN,    "weight", i), {hc * n_embd, 2 * hc}, 0);
+        layer.hc_attn_base  = create_tensor(tn(LLM_TENSOR_HC_ATTN_BASE,  "weight", i), {2 * hc}, 0);
+        layer.hc_attn_scale = create_tensor(tn(LLM_TENSOR_HC_ATTN_SCALE, "weight", i), {2}, 0);
+        layer.hc_ffn_fn     = create_tensor(tn(LLM_TENSOR_HC_FFN_FN,     "weight", i), {hc * n_embd, 2 * hc}, 0);
+        layer.hc_ffn_base   = create_tensor(tn(LLM_TENSOR_HC_FFN_BASE,   "weight", i), {2 * hc}, 0);
+        layer.hc_ffn_scale  = create_tensor(tn(LLM_TENSOR_HC_FFN_SCALE,  "weight", i), {2}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        } else {
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+            if (n_expert == 0) {
+                throw std::runtime_error("n_expert must be > 0");
+            }
+            if (n_expert_used == 0) {
+                throw std::runtime_error("n_expert_used must be > 0");
+            }
+
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd,   n_expert}, 0);
+
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_exp * n_expert_shared, n_embd}, 0);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_hy_v4::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+// reduce hc streams x[:,i,:] weighted by w[i,:] -> [n_embd, n_tokens]
+// reference runs this in fp32 (inside the float() / autocast(fp32) context)
+static ggml_tensor * hy_v4_hc_reduce(ggml_context * ctx0, ggml_tensor * x, ggml_tensor * w, int64_t hc, int64_t n_embd, int64_t nt, ggml_type out_type) {
+    ggml_tensor * x_f32 = ggml_cast(ctx0, x, GGML_TYPE_F32);
+    ggml_tensor * result = nullptr;
+    for (int64_t ih = 0; ih < hc; ++ih) {
+        ggml_tensor * xh = ggml_view_2d(ctx0, x_f32, n_embd, nt, x_f32->nb[2], ih * x_f32->nb[1]);
+        ggml_tensor * wh = ggml_view_2d(ctx0, w, 1, nt, w->nb[1], ih * w->nb[0]);
+        ggml_tensor * cur = ggml_mul(ctx0, xh, wh);
+        result = result ? ggml_add(ctx0, result, cur) : cur;
+    }
+    return ggml_cast(ctx0, result, out_type);
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_hc_pre(
+        ggml_tensor * x,
+        ggml_tensor * hc_fn,
+        ggml_tensor * hc_scale,
+        ggml_tensor * hc_base,
+        ggml_tensor ** post,
+        int il) const {
+    const int64_t hc  = hparams.dsv4_hc_mult;
+    const int64_t nt  = x->ne[2];
+    GGML_ASSERT(x->ne[0] == n_embd && x->ne[1] == hc);
+
+    ggml_tensor * flat = ggml_reshape_2d(ctx0, x, hc * n_embd, nt);
+    ggml_tensor * flat_norm = ggml_rms_norm(ctx0, flat, hparams.f_norm_rms_eps);
+    ggml_tensor * mixes = ggml_mul_mat(ctx0, hc_fn, flat_norm); // [2*hc, nt]
+    cb(mixes, "hc_mixes", il);
+
+    ggml_tensor * scale_pre  = hy_v4_view_1d(ctx0, hc_scale, 1, 0);
+    ggml_tensor * scale_post = hy_v4_view_1d(ctx0, hc_scale, 1, 1);
+    ggml_tensor * base_pre   = hy_v4_view_1d(ctx0, hc_base, hc, 0);
+    ggml_tensor * base_post  = hy_v4_view_1d(ctx0, hc_base, hc, hc);
+
+    // pre = sigmoid(mixes[:hc]*scale_pre + base_pre) + eps
+    ggml_tensor * pre = hy_v4_view_2d(ctx0, mixes, hc, nt, 0);
+    pre = ggml_mul(ctx0, pre, scale_pre);
+    pre = ggml_add(ctx0, pre, base_pre);
+    pre = ggml_sigmoid(ctx0, pre);
+    pre = ggml_scale_bias(ctx0, pre, 1.0f, hparams.dsv4_hc_eps);
+    cb(pre, "hc_pre", il);
+
+    // post = magnitude*sigmoid(mixes[hc:2hc]*scale_post + base_post) + eps
+    ggml_tensor * po = hy_v4_view_2d(ctx0, mixes, hc, nt, hc);
+    po = ggml_mul(ctx0, po, scale_post);
+    po = ggml_add(ctx0, po, base_post);
+    po = ggml_sigmoid(ctx0, po);
+    po = ggml_scale(ctx0, po, hparams.hc_magnitude);
+    po = ggml_scale_bias(ctx0, po, 1.0f, hparams.dsv4_hc_eps);
+    *post = po;
+    cb(po, "hc_post_gate", il);
+
+    return hy_v4_hc_reduce(ctx0, x, pre, hc, n_embd, nt, x->type);
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_hc_post(
+        ggml_tensor * x,
+        ggml_tensor * residual,
+        ggml_tensor * post,
+        int il) const {
+    GGML_UNUSED(il);
+    const int64_t hc = hparams.dsv4_hc_mult;
+    const int64_t nt = x->ne[1];
+    GGML_ASSERT(x->ne[0] == n_embd);
+    GGML_ASSERT(residual->ne[1] == hc);
+
+    // reference HC post runs entirely in fp32 to avoid bf16 rounding accumulation
+    // across 78 layers: post.float() * x.float() + residual.float() -> .to(dtype)
+    ggml_tensor * x_f32    = ggml_cast(ctx0, x, GGML_TYPE_F32);
+    ggml_tensor * post_f32 = ggml_cast(ctx0, post, GGML_TYPE_F32);
+    ggml_tensor * res_f32  = ggml_cast(ctx0, residual, GGML_TYPE_F32);
+
+    ggml_tensor * out = nullptr;
+    for (int64_t i = 0; i < hc; ++i) {
+        ggml_tensor * res_i  = ggml_view_2d(ctx0, res_f32, n_embd, nt, res_f32->nb[2], i * res_f32->nb[1]);
+        ggml_tensor * post_i = ggml_view_2d(ctx0, post_f32, 1, nt, post_f32->nb[1], i * post_f32->nb[0]);
+        ggml_tensor * cur = ggml_add(ctx0, res_i, ggml_mul(ctx0, x_f32, post_i));
+        cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, nt);
+        out = out ? ggml_concat(ctx0, out, cur, 1) : cur;
+    }
+
+    // cast back to the original type (bf16)
+    out = ggml_cast(ctx0, out, residual->type);
+    return out; // [n_embd, hc, nt]
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_hc_head(
+        ggml_tensor * x,
+        ggml_tensor * hc_fn,
+        ggml_tensor * hc_scale,
+        ggml_tensor * hc_base) const {
+    const int64_t hc = hparams.dsv4_hc_mult;
+    const int64_t nt = x->ne[2];
+
+    ggml_tensor * flat = ggml_reshape_2d(ctx0, x, hc * n_embd, nt);
+    ggml_tensor * flat_norm = ggml_rms_norm(ctx0, flat, hparams.f_norm_rms_eps);
+    ggml_tensor * mixes = ggml_mul_mat(ctx0, hc_fn, flat_norm); // [hc, nt]
+    cb(mixes, "hc_head_mixes", -1);
+
+    ggml_tensor * pre = ggml_mul(ctx0, mixes, hc_scale);
+    pre = ggml_add(ctx0, pre, hc_base);
+    pre = ggml_sigmoid(ctx0, pre);
+    pre = ggml_scale_bias(ctx0, pre, 1.0f, hparams.dsv4_hc_eps);
+    cb(pre, "hc_head_pre", -1);
+
+    return hy_v4_hc_reduce(ctx0, x, pre, hc, n_embd, nt, x->type);
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_attention(
+        const llama_model & model,
+        llm_graph_input_attn_k * inp_attn,
+        ggml_tensor * cur,
+        ggml_tensor * inp_pos,
+        float kq_scale,
+        int il) const {
+    const auto & layer = model.layers[il];
+
+    const int64_t n_embd_head_k       = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+    const uint32_t kv_lora_rank       = hparams.n_lora_kv;
+
+    ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);
+    q = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+    q = ggml_mul_mat(ctx0, layer.wq_b, q);
+
+    ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+            ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+    ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+            ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head,
+            ggml_row_size(q->type, n_embd_head_qk_nope));
+
+    ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+    ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+    ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+
+    q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    cb(q_pe, "q_pe", il);
+    k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    cb(k_pe, "k_pe", il);
+
+    kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+    cb(kv_cmpr, "kv_cmpr", il);
+
+    // MLA absorption: q_nope @ wk_b -> compressed space
+    q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+    ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+    q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+
+    // note: rope must go first for in-place context shifting in build_rope_shift()
+    ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+
+    kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+    ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+    ggml_tensor * Vcur = kv_cmpr;
+
+    // MLA-as-MQA; wo applied manually below so the gated-MLA gate can sit before o_proj
+    ggml_tensor * attn = build_attn(inp_attn,
+            nullptr, nullptr, nullptr,
+            Qcur, Kcur, Vcur, nullptr, layer.attn_sinks, layer.wv_b, kq_scale, il);
+    cb(attn, "attn_kqv", il); // [n_head * n_embd_head_v, n_tokens]
+
+    // gated MLA: elementwise sigmoid gate on the decompressed attention output
+    ggml_tensor * gate = ggml_mul_mat(ctx0, layer.wqkv_gate, cur);
+    gate = ggml_sigmoid(ctx0, gate);
+    attn = ggml_mul(ctx0, attn, gate);
+    cb(attn, "attn_gated", il);
+
+    ggml_tensor * out = build_lora_mm(layer.wo, attn);
+    cb(out, "attn_out", il);
+
+    return out;
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_indexer_top_k(
+        const llama_model & model,
+        llm_graph_input_attn_k_dsa * inp_attn_dsa,
+        ggml_tensor * cur,
+        ggml_tensor * qr,
+        ggml_tensor * inp_pos,
+        int il) const {
+    const auto & layer = model.layers[il];
+
+    const int64_t n_indexer_head       = hparams.indexer_n_head;
+    const int64_t n_embd_indexer       = hparams.indexer_head_size;
+    const int64_t n_embd_indexer_rope  = hparams.n_rot();
+    const int64_t n_embd_indexer_nope  = n_embd_indexer - n_embd_indexer_rope;
+
+    // nope rows come first, so rope only the last n_embd_indexer_rope rows, same as the MLA path
+    ggml_tensor * iq = ggml_mul_mat(ctx0, layer.indexer_attn_q_b, qr);
+
+    iq = ggml_reshape_3d(ctx0, iq, n_embd_indexer, n_indexer_head, n_tokens);
+
+    iq = ggml_rope_ext(ctx0, iq, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base,
+         freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+    iq = ggml_rope_set_offset(iq, n_embd_indexer_nope);
+    cb(iq, "indexer_q", il);
+
+    ggml_tensor * ik = ggml_mul_mat(ctx0, layer.indexer_attn_k, cur);
+
+    ik = build_norm(ik, layer.indexer_k_norm, layer.indexer_k_norm_b, LLM_NORM, il);
+
+    ik = ggml_reshape_3d(ctx0, ik, n_embd_indexer, 1, n_tokens);
+
+    ik = ggml_rope_ext(ctx0, ik, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base,
+         freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+    ik = ggml_rope_set_offset(ik, n_embd_indexer_nope);
+    cb(ik, "indexer_k", il);
+
+    // the reference applies a Hadamard rotation here, but it only helps its FP8 kernels.
+    // it is orthogonal, so it does not change q.k and we can skip it.
+
+    const auto * mctx_lid   = inp_attn_dsa->mctx->get_lid();
+    const auto & k_idxs_lid = inp_attn_dsa->get_k_idxs_lid();
+    ggml_build_forward_expand(gf, mctx_lid->cpy_k(ctx0, ik, k_idxs_lid, il));
+
+    ggml_tensor * iw = ggml_mul_mat(ctx0, layer.indexer_proj, cur);
+
+    ik = mctx_lid->get_k(ctx0, il);
+
+    const auto n_stream = ik->ne[3];
+    iq = ggml_view_4d(ctx0, iq, iq->ne[0], iq->ne[1], iq->ne[2]/n_stream, n_stream,
+            iq->nb[1], iq->nb[2], iq->nb[3]/n_stream, 0);
+    iw = ggml_view_4d(ctx0, iw, iw->ne[0], iw->ne[1]/n_stream, iw->ne[2], n_stream,
+            iw->nb[1], iw->nb[2]/n_stream, iw->nb[3]/n_stream, 0);
+
+    // fold both reference scale factors into the weights before the big score tensor
+    iw = ggml_scale(ctx0, iw, 1.0f / sqrtf(float(n_embd_indexer * n_indexer_head)));
+
+    ggml_tensor * score = nullptr;
+    if (cparams.fused_lid) {
+        score = ggml_lightning_indexer(ctx0, iq, ik, iw, inp_attn_dsa->get_kq_mask_lid());
+        cb(score, "indexer_score", il);
+        res->add_fused_node({LLM_FUSED_OP_LIGHTNING_INDEXER, score, il});
+    } else {
+        iq = ggml_permute(ctx0, iq, 0, 2, 1, 3);
+        ik = ggml_permute(ctx0, ik, 0, 2, 1, 3);
+
+        score = ggml_mul_mat(ctx0, ik, iq);
+        score = ggml_cont(ctx0, ggml_permute(ctx0, score, 2, 1, 0, 3));
+        score = ggml_relu(ctx0, score);
+        score = ggml_mul(ctx0, score, iw);
+        score = ggml_sum_rows(ctx0, score);
+        score = ggml_cont(ctx0, ggml_permute(ctx0, score, 2, 1, 0, 3));
+        score = ggml_add(ctx0, score, inp_attn_dsa->get_kq_mask_lid());
+        cb(score, "indexer_score", il);
+    }
+
+    const uint32_t n_top_k = score->ne[0] < (int64_t) hparams.indexer_top_k ? score->ne[0] : hparams.indexer_top_k;
+
+    return ggml_cont(ctx0, ggml_top_k(ctx0, score, n_top_k));
+}
+
+ggml_tensor * llama_model_hy_v4::graph::build_attention_dsa(
+        const llama_model & model,
+        llm_graph_input_attn_k_dsa * inp_attn_dsa,
+        ggml_tensor * cur,
+        ggml_tensor * inp_pos,
+        ggml_tensor ** last_top_k,
+        float kq_scale,
+        int il) const {
+    const auto & layer = model.layers[il];
+
+    const int64_t n_embd_head_k       = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+    const uint32_t kv_lora_rank       = hparams.n_lora_kv;
+
+    ggml_tensor * qr = ggml_mul_mat(ctx0, layer.wq_a, cur);
+    qr = build_norm(qr, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+
+    if (hparams.is_indexer_full(il)) {
+        *last_top_k = build_indexer_top_k(model, inp_attn_dsa, cur, qr, inp_pos, il);
+        cb(*last_top_k, "top_k", il);
+    }
+    GGML_ASSERT(*last_top_k != nullptr);
+
+    ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_b, qr);
+
+    ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+            ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+    ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+            ggml_row_size(q->type, n_embd_head_k), ggml_row_size(q->type, n_embd_head_k) * n_head,
+            ggml_row_size(q->type, n_embd_head_qk_nope));
+
+    ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+    ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+    ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+            ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+
+    q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    cb(q_pe, "q_pe", il);
+    k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    cb(k_pe, "k_pe", il);
+
+    kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+    cb(kv_cmpr, "kv_cmpr", il);
+
+    q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+    ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+    q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+
+    ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+
+    kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+    ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+    ggml_tensor * Vcur = kv_cmpr;
+
+    ggml_tensor * attn = build_attn(inp_attn_dsa,
+            nullptr, nullptr, nullptr,
+            Qcur, Kcur, Vcur, nullptr, layer.attn_sinks, layer.wv_b, *last_top_k, kq_scale, il);
+    cb(attn, "attn_kqv", il);
+
+    ggml_tensor * gate = ggml_mul_mat(ctx0, layer.wqkv_gate, cur);
+    gate = ggml_sigmoid(ctx0, gate);
+    attn = ggml_mul(ctx0, attn, gate);
+    cb(attn, "attn_gated", il);
+
+    ggml_tensor * out = build_lora_mm(layer.wo, attn);
+    cb(out, "attn_out", il);
+
+    return out;
+}
+
+llama_model_hy_v4::graph::graph(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    const int64_t hc = hparams.dsv4_hc_mult;
+    const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+    const float kq_scale = 1.0f / sqrtf(float(n_embd_head_k));
+
+    ggml_tensor * cur;
+
+    const bool is_dsa = hparams.indexer_top_k > 0;
+
+    ggml_tensor * inp = build_inp_embd(model.tok_embd);
+    ggml_tensor * inp_pos = build_inp_pos();
+    llm_graph_input_attn_k     * inp_attn     = is_dsa ? nullptr : build_attn_inp_k();
+    llm_graph_input_attn_k_dsa * inp_attn_dsa = is_dsa ? build_attn_inp_k_dsa() : nullptr;
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // top-k of the last "full" indexer layer, reused by the following "shared" layers
+    ggml_tensor * last_top_k = nullptr;
+
+    // expand the single embedding into hc parallel residual streams
+    ggml_tensor * inpL = ggml_reshape_3d(ctx0, inp, n_embd, 1, n_tokens);
+    inpL = ggml_repeat_4d(ctx0, inpL, n_embd, hc, n_tokens, 1);
+    cb(inpL, "hc_init", -1);
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * residual = inpL;
+        ggml_tensor * post = nullptr;
+
+        cur = build_hc_pre(inpL, model.layers[il].hc_attn_fn, model.layers[il].hc_attn_scale,
+                model.layers[il].hc_attn_base, &post, il);
+        cur = build_norm(cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        cur = is_dsa
+            ? build_attention_dsa(model, inp_attn_dsa, cur, inp_pos, &last_top_k, kq_scale, il)
+            : build_attention(model, inp_attn, cur, inp_pos, kq_scale, il);
+
+        inpL = build_hc_post(cur, residual, post, il);
+        cb(inpL, "hc_attn_out", il);
+
+        residual = inpL;
+        cur = build_hc_pre(inpL, model.layers[il].hc_ffn_fn, model.layers[il].hc_ffn_scale,
+                model.layers[il].hc_ffn_base, &post, il);
+        cur = build_norm(cur, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        const auto & layer = model.layers[il];
+        if ((uint32_t) il < hparams.n_layer_dense_lead) {
+            cur = build_ffn(cur,
+                    layer.ffn_up,   NULL, NULL,
+                    layer.ffn_gate, NULL, NULL,
+                    layer.ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            ggml_tensor * moe_out = build_moe_ffn(cur,
+                    layer.ffn_gate_inp,
+                    layer.ffn_up_exps,
+                    layer.ffn_gate_exps,
+                    layer.ffn_down_exps,
+                    layer.ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, hparams.expert_weights_norm,
+                    hparams.expert_weights_scale,
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il,
+                    nullptr,
+                    nullptr);
+            cb(moe_out, "ffn_moe_out", il);
+
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                    layer.ffn_up_shexp,   NULL, NULL,
+                    layer.ffn_gate_shexp, NULL, NULL,
+                    layer.ffn_down_shexp, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        }
+
+        inpL = build_hc_post(cur, residual, post, il);
+        cb(inpL, "l_out", il);
+    }
+
+    // prune to the requested output rows once, after all HC streams are done
+    if (inp_out_ids) {
+        ggml_tensor * flat = ggml_reshape_2d(ctx0, inpL, n_embd * hc, n_tokens);
+        flat = ggml_get_rows(ctx0, flat, inp_out_ids);
+        inpL = ggml_reshape_3d(ctx0, flat, n_embd, hc, n_outputs);
+    }
+
+    cur = build_hc_head(inpL, model.hc_head_fn, model.hc_head_scale, model.hc_head_base);
+    cb(cur, "hc_head", -1);
+
+    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    cur = ggml_mul_mat(ctx0, model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}
@@ -29,15 +29,9 @@ void llama_model_jais2::load_arch_tensors(llama_model_loader &) {
        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);

-        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
-        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
-        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);

-        // attention biases - all have shape n_embd (output dimension of projections)
-        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);
-        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd}, 0);
-        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd}, 0);
        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);

        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@@ -441,9 +441,9 @@ ggml_tensor * llama_model_kimi_k3::graph::build_kda_layer(
    ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
    state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head_kda, n_seqs);

-    const float eps = hparams.f_norm_rms_eps;
-    Qcur = ggml_l2_norm(ctx0, Qcur, eps);
-    Kcur = ggml_l2_norm(ctx0, Kcur, eps);
+    const float eps_norm = hparams.f_norm_rms_eps;
+    Qcur = build_gdn_l2_norm(ctx0, Qcur, eps_norm);
+    Kcur = build_gdn_l2_norm(ctx0, Kcur, eps_norm);

    auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);

@@ -195,7 +195,7 @@ static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_t
 // Causal Conv1d function for Q,K,V
 // When qkv is 0, it is Q, 1 is K, 2 is V
    // Step 1: Q, K, V projections -> [d_inner, n_tokens]
-    ggml_tensor * x_proj = ggml_mul_mat(ctx0, proj_w, x);
+    ggml_tensor * x_proj = proj_w ? ggml_mul_mat(ctx0, proj_w, x) : x;

    // Reshape input: {d_inner, n_tokens} -> {d_inner, n_seq_tokens, n_seqs}
    ggml_tensor * x_3d = ggml_reshape_3d(ctx0, x_proj, d_inner, n_seq_tokens, n_seqs);
@@ -295,9 +295,20 @@ llama_model_kimi_linear::graph::graph(const llama_model & model, const llm_graph
            ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
            cb(conv_states_all, "conv_states_all", il);
            ggml_tensor * conv_state_all = build_rs(inp_rs, conv_states_all, hparams.n_embd_r(), n_seqs);
-            ggml_tensor * Qcur = causal_conv1d(gf, ctx0, conv_states_all, conv_state_all, 0, cur, layer.wq, layer.ssm_q_conv, d_conv, head_dim, n_head, n_seq_tokens, n_seqs, n_tokens, kv_head);
-            ggml_tensor * Kcur = causal_conv1d(gf, ctx0, conv_states_all, conv_state_all, 1, cur, layer.wk, layer.ssm_k_conv, d_conv, head_dim, n_head, n_seq_tokens, n_seqs, n_tokens, kv_head);
-            ggml_tensor * Vcur = causal_conv1d(gf, ctx0, conv_states_all, conv_state_all, 2, cur, layer.wv, layer.ssm_v_conv, d_conv, head_dim, n_head, n_seq_tokens, n_seqs, n_tokens, kv_head);
+            ggml_tensor * q_in = cur, * k_in = cur, * v_in = cur;
+            ggml_tensor * q_w = layer.wq, * k_w = layer.wk, * v_w = layer.wv;
+            if (layer.wqkv) {
+                ggml_tensor * qkv = ggml_mul_mat(ctx0, layer.wqkv, cur);
+                const int64_t d_inner = head_dim * n_head;
+                const size_t esize = ggml_element_size(qkv);
+                q_in = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, d_inner, n_tokens, qkv->nb[1], 0));
+                k_in = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, d_inner, n_tokens, qkv->nb[1], d_inner * esize));
+                v_in = ggml_cont(ctx0, ggml_view_2d(ctx0, qkv, d_inner, n_tokens, qkv->nb[1], 2 * d_inner * esize));
+                q_w = nullptr; k_w = nullptr; v_w = nullptr;
+            }
+            ggml_tensor * Qcur = causal_conv1d(gf, ctx0, conv_states_all, conv_state_all, 0, q_in, q_w, layer.ssm_q_conv, d_conv, head_dim, n_head, n_seq_tokens, n_seqs, n_tokens, kv_head);
+            ggml_tensor * Kcur = causal_conv1d(gf, ctx0, conv_states_all, conv_state_all, 1, k_in, k_w, layer.ssm_k_conv, d_conv, head_dim, n_head, n_seq_tokens, n_seqs, n_tokens, kv_head);
+            ggml_tensor * Vcur = causal_conv1d(gf, ctx0, conv_states_all, conv_state_all, 2, v_in, v_w, layer.ssm_v_conv, d_conv, head_dim, n_head, n_seq_tokens, n_seqs, n_tokens, kv_head);

            // g1 = -exp(A_log) * softplus(f_b(f_a(x)) + dt_bias)
            ggml_tensor * f_a = ggml_mul_mat(ctx0, layer.ssm_f_a, cur);
@@ -331,10 +342,11 @@ llama_model_kimi_linear::graph::graph(const llama_model & model, const llm_graph
            ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
            state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head, n_seqs);

+
            const float eps_norm = hparams.f_norm_rms_eps;

-            Qcur = ggml_l2_norm(ctx0, Qcur, eps_norm);
-            Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
+            Qcur = build_gdn_l2_norm(ctx0, Qcur, eps_norm);
+            Kcur = build_gdn_l2_norm(ctx0, Kcur, eps_norm);

            // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
            auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);
--- a/Show More
+++ b/Show More