opencl: write KV-cache views in set_tensor so a state restore is not dropped (#30471 )

hexagon: MUL_MAT_ID row skipping optimizations and improvements (#31169 )
* hex-mmid: vectorize skipped row zeroing and added zeroing to fused MM_ID_NX * hex-mmid: remove scalar div
2026-10-26 14:37:36 -05:00 · 2026-10-26 11:39:22 -07:00 · 2026-10-26 08:21:14 -07:00 · 2026-10-26 16:21:25 +02:00 · 2026-10-26 16:19:47 +02:00 · 2026-10-26 16:18:45 +02:00
497 changed files with 49734 additions and 25980 deletions
@@ -0,0 +1,109 @@
+name: Behavior Check
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/behavior-check.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/behavior-check.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+cache-mode: none
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  GGML_SCHED_DEBUG_REALLOC: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu:
+    runs-on: ubuntu-26.04-arm
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.24
+        with:
+          restore: false
+          save: false
+
+      - name: Build Dependencies
+        id: build_depends
+        run: |
+          sudo apt update
+          sudo apt install -y build-essential libssl-dev git-lfs libjpeg-dev python3 python3-venv python3-pip
+
+      - name: Python Dependencies
+        id: python_depends
+        run: |
+          python3 -m venv .venv
+          source .venv/bin/activate
+          pip install jinja2==3.1.6
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: behavior-check-ubuntu-26.04-arm
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DGGML_NATIVE=OFF \
+            -DGGML_BACKEND_DL=OFF \
+            -DLLAMA_FATAL_WARNINGS=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: behavior-check-ubuntu-26.04-arm
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
+      - name: Test
+        id: cmake_test
+        run: |
+          source .venv/bin/activate
+          cd build
+          ctest -L behavior -LE 'model|backend' --verbose --timeout 900
@@ -105,63 +105,7 @@ jobs:
          cd build
          # Metal Paravirtual devices are difficult to support -> disable
          # ref: https://github.com/ggml-org/llama.cpp/pull/19802#issuecomment-4013704023
-          ctest -L main -E "test-llama-archs|test-llama-context" --verbose --timeout 900
-
-  macos-latest-x64:
-    runs-on: macos-15-intel
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          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
-        run: |
-          sysctl -a
-          # Metal is disabled due to intermittent failures with Github runners not having a GPU:
-          # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_METAL=OFF \
-            -DGGML_RPC=ON \
-            -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
+          ctest -L main -E "test-llama-archs|test-llama-context|test-speculative-rejection-sampling" --verbose --timeout 900

  macos-latest-ios-xcode:
    runs-on: macos-latest
@@ -9,7 +9,7 @@ permissions:

 jobs:
  linux:
-    runs-on: [self-hosted, Linux, CPU]
+    runs-on: [self-hosted, Linux, fast]
    steps:
      - uses: actions/checkout@v6
        with:
@@ -129,7 +129,7 @@ jobs:
        id: cmake_test
        run: |
          cd build
-          ctest -L 'main|python' --verbose --timeout 900
+          ctest -L main --verbose --timeout 900

      - name: Test llama2c conversion
        id: llama2c_test
@@ -49,7 +49,7 @@ jobs:
        - sanitizer: THREAD
          machine: ubuntu-24.04
        - sanitizer: UNDEFINED
-          machine: [self-hosted, X64, Linux]
+          machine: [self-hosted, X64, Linux, CPU]

    runs-on: ${{ matrix.machine }}

@@ -7,15 +7,7 @@ on:
      - master
    paths: [
      '.github/workflows/build-webgpu.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp',
-      '**/*.wgsl',
-      '**/*.tmpl',
-      'ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py'
+      'ggml/src/ggml-webgpu/**'
    ]

  pull_request:
@@ -63,135 +55,3 @@ jobs:
            -type f \( -name '*.cpp' -o -name '*.hpp' -o -name '*.h' \) \
            -print0 |
            xargs -0 clang-format-22 --dry-run --Werror
-
-  macos:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          DAWN_VERSION="v20260908.214631"
-          DAWN_OWNER="google"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-macos-latest-Release"
-          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          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: |
-          export CMAKE_PREFIX_PATH=dawn
-          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: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: webgpu-macos-latest
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-  ubuntu:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo add-apt-repository -y ppa:kisak/kisak-mesa
-          sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers \
-            libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libssl-dev
-
-      - name: Dawn Dependency
-        id: dawn-depends
-        run: |
-          sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          DAWN_VERSION="v20260908.214631"
-          DAWN_OWNER="google"
-          DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-94c3c9cc0d5fb2e85aebb370fa8d37b71aa34655-ubuntu-latest-Release"
-          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          curl -L -o artifact.tar.gz \
-            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
-          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: |
-          export Dawn_DIR=dawn/lib64/cmake/Dawn
-          cmake -B build \
-            -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: |
-          cd build
-          ctest -L main --verbose --timeout 900
@@ -46,7 +46,7 @@ env:

 jobs:
  cpu-x64-high-perf:
-    runs-on: [self-hosted, Linux, X64]
+    runs-on: [self-hosted, Linux, X64, CPU]

    steps:
      - name: Clone
@@ -56,7 +56,7 @@ jobs:
      - name: Test
        id: ggml-ci
        run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp/$RUNNER_NAME ~/mnt/llama.cpp

  cpu-arm64-high-perf-graviton4:
    runs-on: ah-ubuntu_24_04-c8g_8x
@@ -60,4 +60,4 @@ jobs:
      - name: Test
        id: ggml-ci
        run: |
-          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp/$RUNNER_NAME ~/mnt/llama.cpp
@@ -79,7 +79,7 @@ jobs:
        id: ggml-ci
        run: |
          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp/$RUNNER_NAME ~/mnt/llama.cpp

      # - name: ccache-buckets-save
      #   if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -124,7 +124,7 @@ jobs:
        id: ggml-ci
        run: |
          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp/$RUNNER_NAME ~/mnt/llama.cpp

      # - name: ccache-buckets-save
      #   if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -150,7 +150,7 @@ jobs:
        id: ggml-ci
        run: |
          vulkaninfo --summary
-          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp/$RUNNER_NAME ~/mnt/llama.cpp

  gpu-vulkan-intel-linux:
    runs-on: [self-hosted, Linux, Intel]
@@ -131,4 +131,4 @@ jobs:
        id: ggml-ci
        run: |
          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
-            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+            bash ./ci/run.sh ~/results/llama.cpp/$RUNNER_NAME ~/mnt/llama.cpp
@@ -185,18 +185,28 @@ fi

 ## helpers

-# download a file if it does not exist or if it is outdated
+# download a file unless already cached in $MNT
 function gg_wget {
    local out=$1
    local url=$2

    local cwd=`pwd`
+    local name=`basename $url`
+    local tmp=.gg-wget.$$.tmp

    mkdir -p $out
    cd $out

-    # should not re-download if file is the same
-    wget -nv -c -N $url
+    # cache first: HF artifacts are static, HF sends no Last-Modified for -N
+    if [ ! -s $name ]; then
+        # publish with mv: concurrent runners never read partial files
+        rm -rf $tmp && mkdir $tmp
+        (cd $tmp && wget -nv $url)
+        if [ -s $tmp/$name ]; then
+            mv -f $tmp/$name $name
+        fi
+        rm -rf $tmp
+    fi

    cd $cwd
 }
@@ -254,7 +264,7 @@ function gg_run_ctest_release {
    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log

    if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest -C Release --output-on-failure -L 'main|python' ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest -C Release --output-on-failure -L main ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
    else
        (time ctest -C Release --output-on-failure -L main -E test-opt ${CTEST_EXTRA}) 2>&1 | tee -a $OUT/${ci}-ctest.log
    fi
@@ -312,8 +322,8 @@ function gg_run_test_scripts {

    set -e

-    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$OUT/scratch") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$OUT/scratch") 2>&1 | tee -a $OUT/${ci}-scripts.log

    set +e
 }
@@ -374,10 +384,14 @@ function gg_run_qwen3_0_6b {


    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    rm -rf $OUT/scratch/qwen3-0.6B $OUT/scratch/wikitext
+    mkdir -p $OUT/scratch/qwen3-0.6B
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d $OUT/scratch/wikitext/

    path_models="../models-mnt/qwen3/0.6B"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+    path_local="$OUT/scratch/qwen3-0.6B"
+    path_wiki="$OUT/scratch/wikitext/wikitext-2-raw"

    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release

@@ -386,72 +400,15 @@ function gg_run_qwen3_0_6b {
    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log

-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf  --outtype f16
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-bf16.gguf --outtype bf16
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_local}/ggml-model-f16.gguf  --outtype f16
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_local}/ggml-model-bf16.gguf --outtype bf16

-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_bf16="${path_models}/ggml-model-bf16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+    model_f16="${path_local}/ggml-model-f16.gguf"
+    model_bf16="${path_local}/ggml-model-bf16.gguf"
+    model_q4_0="${path_local}/ggml-model-q4_0.gguf"

    wiki_test="${path_wiki}/wiki.test.raw"

-    ./bin/llama-quantize ${model_bf16} ${model_q8_0} q8_0 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q4_0} q4_0 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q4_1} q4_1 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q5_0} q5_0 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q5_1} q5_1 $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q2_k} q2_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q3_k} q3_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q4_k} q4_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k $(nproc)
-    ./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k $(nproc)
-
-    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
-
-    (time ./bin/llama-completion -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-completion -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q4_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q4_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q5_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q5_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q2_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q3_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q4_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q5_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-completion -no-cnv --model ${model_q6_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    if [ -z ${GG_BUILD_NO_BF16} ]; then
-        (time ./bin/llama-perplexity --model ${model_bf16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
-    fi
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/test-llama-context --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
-    (time ./bin/test-llama-context --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
-    (time ./bin/test-llama-context --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
-    (time ./bin/test-llama-context --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
-
    function check_ppl {
        qnt="$1"
        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@@ -465,20 +422,56 @@ function gg_run_qwen3_0_6b {
        return 0
    }

+    for qnt in q8_0 q4_0 q4_1 q5_0 q5_1 q2_k q3_k q4_k q5_k q6_k; do
+        model="${path_local}/ggml-model-${qnt}.gguf"
+
+        ./bin/llama-quantize ${model_bf16} ${model} ${qnt} $(nproc)
+
+        (time ./bin/llama-completion -no-cnv --model ${model} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-${qnt}.log
+        (time ./bin/llama-perplexity --model ${model} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-${qnt}.log
+
+        # note: ppl > 20.0 for q2_k with this model
+        if [ "${qnt}" != "q2_k" ]; then
+            check_ppl "${qnt}" "$(cat $OUT/${ci}-tg-${qnt}.log | grep "^\[1\]")"
+        fi
+
+        if [ "${qnt}" = "q4_0" ]; then
+            # q4_0 is kept for other tests
+            q4_0="${path_models}/ggml-model-q4_0.gguf"
+            ln ${model} "$q4_0" 2>/dev/null || { cp ${model} $q4_0.tmp.$$ && mv -f $q4_0.tmp.$$ $q4_0; }
+        else
+            rm -f ${model}
+        fi
+    done
+
+    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)
+
+    (time ./bin/llama-completion -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-completion -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
+
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    if [ -z ${GG_BUILD_NO_BF16} ]; then
+        (time ./bin/llama-perplexity --model ${model_bf16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
+    fi
+
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    # TODO: these tests are extremely slow on WebGPU - re-enable when fixed
+    #       https://github.com/ggml-org/llama.cpp/actions/runs/39639985732
+    if [ -z ${GG_BUILD_WEBGPU} ]; then
+        (time ./bin/test-llama-context --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
+        (time ./bin/test-llama-context --model ${model_q4_0} -ngl 99 -c 1024 -fa off                ) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
+    fi
+    (time ./bin/test-llama-context --model ${model_q4_0} -ngl 10 -c 1024 -fa on  --no-op-offload) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
+    (time ./bin/test-llama-context --model ${model_q4_0} -ngl 99 -c 1024 -fa on                 ) 2>&1 | tee -a $OUT/${ci}-test-llama-context.log
+
    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")"
    if [ -z ${GG_BUILD_NO_BF16} ]; then
        check_ppl "bf16" "$(cat $OUT/${ci}-tg-bf16.log | grep "^\[1\]")"
    fi
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")"
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")"
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")"
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")"
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")"
-   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" # note: ppl > 20.0 for this quant and model
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")"
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")"
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")"
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")"
+
+    # scratch is rebuilt on every run; q4_0 is already published to $MNT
+    rm -rf $OUT/scratch/qwen3-0.6B $OUT/scratch/wikitext

    set +e
 }
@@ -501,6 +494,8 @@ function gg_run_embd_bge_small {
    gg_wget models-mnt/bge-small/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json

    path_models="../models-mnt/bge-small"
+    path_local="$OUT/scratch/bge-small"
+    rm -rf $path_local && mkdir -p $path_local

    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release

@@ -509,10 +504,10 @@ function gg_run_embd_bge_small {
    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log

-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_local}/ggml-model-f16.gguf

-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_f16="${path_local}/ggml-model-f16.gguf"
+    model_q8_0="${path_local}/ggml-model-q8_0.gguf"

    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0

@@ -521,6 +516,8 @@ function gg_run_embd_bge_small {
    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log

+    rm -rf $path_local
+
    set +e
 }

@@ -537,6 +534,8 @@ function gg_run_rerank_tiny {
    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.json

    path_models="../models-mnt/rerank-tiny"
+    path_local="$OUT/scratch/rerank-tiny"
+    rm -rf $path_local && mkdir -p $path_local

    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release

@@ -545,9 +544,9 @@ function gg_run_rerank_tiny {
    (cmake -G "${CMAKE_GENERATOR}" -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time cmake --build . --config Release -j$(nproc)) 2>&1 | tee -a $OUT/${ci}-make.log

-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_local}/ggml-model-f16.gguf

-    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_f16="${path_local}/ggml-model-f16.gguf"

    (time ./bin/llama-fit-params --model ${model_f16} 2>&1 | tee -a $OUT/${ci}-fp-f16.log)

@@ -577,6 +576,8 @@ function gg_run_rerank_tiny {
    check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.30" | tee -a $OUT/${ci}-rk-f16.log

+    rm -rf $path_local
+
    set +e
 }

@@ -683,11 +684,11 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
    ln -sfn ${mnt_models} ${SRC}/models-mnt

    # Create a fresh python3 venv and enter it
-    if ! python3 -m venv "$MNT/venv"; then
-        echo "Error: Failed to create Python virtual environment at $MNT/venv."
+    if ! python3 -m venv "$OUT/venv"; then
+        echo "Error: Failed to create Python virtual environment at $OUT/venv."
        exit 1
    fi
-    source "$MNT/venv/bin/activate"
+    source "$OUT/venv/bin/activate"

    pip install -r ${SRC}/requirements.txt --disable-pip-version-check
    pip install --editable gguf-py --disable-pip-version-check
@@ -78,6 +78,8 @@ add_library(${TARGET}
    download.h
    fit.cpp
    fit.h
+    file-lock.cpp
+    file-lock.h
    hf-cache.cpp
    hf-cache.h
    http.h
@@ -184,3 +186,8 @@ if (LLAMA_LLGUIDANCE)
 endif()

 target_link_libraries(${TARGET} PUBLIC llama Threads::Threads)
+
+if (CMAKE_SYSTEM_NAME MATCHES "AIX")
+    # AIX's flock() function comes from libbsd.a
+    target_link_libraries(${TARGET} PRIVATE -lbsd)
+endif()
@@ -470,6 +470,14 @@ void common_models_handler_apply(common_models_handler & handler, common_params
    auto opts = handler.opts; // copy
    opts.callback = callback;

+    // another process may hold the cache lock during finalization:
+    // let the download callback cancel that wait too
+    auto finalize = [callback](const hf_cache::hf_file & f) {
+        return hf_cache::finalize_file(f, [callback]() {
+            return !(callback && callback->is_cancelled());
+        });
+    };
+
    // handle plain "url" if needed
    auto handle_url = [&](common_params_model & model) {
        if (!model.url.empty()) {
@@ -518,7 +526,7 @@ void common_models_handler_apply(common_models_handler & handler, common_params
    }

    // handle hf_plan tasks
-    auto add_tasks = [&opts, &tasks](const hf_cache::hf_files  & model_files,
+    auto add_tasks = [&opts, &tasks, &finalize](const hf_cache::hf_files  & model_files,
                                    const hf_cache::hf_file    & primary,
                                    common_params_model        & model) {
        for (size_t i = 0; i < model_files.size(); ++i) {
@@ -527,9 +535,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
            tasks.emplace_back(model_file, opts, [&, is_primary]() {
                if (is_primary) {
                    // the primary file is the first split (00001-of), use it as model path
-                    model.path = hf_cache::finalize_file(model_file);
+                    model.path = finalize(model_file);
                } else {
-                    hf_cache::finalize_file(model_file);
+                    finalize(model_file);
                }
            });
        }
@@ -581,9 +589,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan_spec.mtp, opts, [&]() {
            // only use the discovered MTP head when no draft path is set yet
            if (params.speculative.draft.mparams.path.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.mtp);
+                params.speculative.draft.mparams.path = finalize(plan_spec.mtp);
            } else {
-                hf_cache::finalize_file(plan_spec.mtp);
+                finalize(plan_spec.mtp);
            }
        });
    }
@@ -591,9 +599,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan_spec.dflash, opts, [&]() {
            // only use the discovered DFlash sidecar when no draft path is set yet
            if (params.speculative.draft.mparams.path.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.dflash);
+                params.speculative.draft.mparams.path = finalize(plan_spec.dflash);
            } else {
-                hf_cache::finalize_file(plan_spec.dflash);
+                finalize(plan_spec.dflash);
            }
        });
    }
@@ -601,9 +609,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan_spec.eagle3, opts, [&]() {
            // only use the discovered Eagle3 sidecar when no draft path is set yet
            if (params.speculative.draft.mparams.path.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.eagle3);
+                params.speculative.draft.mparams.path = finalize(plan_spec.eagle3);
            } else {
-                hf_cache::finalize_file(plan_spec.eagle3);
+                finalize(plan_spec.eagle3);
            }
        });
    }
@@ -611,9 +619,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan_spec.dspark, opts, [&]() {
            // only use the discovered DSpark sidecar when no draft path is set yet
            if (params.speculative.draft.mparams.path.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.dspark);
+                params.speculative.draft.mparams.path = finalize(plan_spec.dspark);
            } else {
-                hf_cache::finalize_file(plan_spec.dspark);
+                finalize(plan_spec.dspark);
            }
        });
    }
@@ -634,16 +642,16 @@ void common_models_handler_apply(common_models_handler & handler, common_params
    }
    if (!plan.mmproj.local_path.empty()) {
        tasks.emplace_back(plan.mmproj, opts, [&]() {
-            params.mmproj.path = hf_cache::finalize_file(plan.mmproj);
+            params.mmproj.path = finalize(plan.mmproj);
        });
    }
    if (!plan.mtp.local_path.empty() && !had_spec_url) {
        tasks.emplace_back(plan.mtp, opts, [&]() {
            // only fall back to the discovered MTP head when no draft was explicitly provided
            if (params.speculative.draft.mparams.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.mtp);
+                params.speculative.draft.mparams.path = finalize(plan.mtp);
            } else {
-                hf_cache::finalize_file(plan.mtp);
+                finalize(plan.mtp);
            }
        });
    }
@@ -651,9 +659,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan.dflash, opts, [&]() {
            // only fall back to the discovered DFlash sidecar when no draft was explicitly provided
            if (params.speculative.draft.mparams.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.dflash);
+                params.speculative.draft.mparams.path = finalize(plan.dflash);
            } else {
-                hf_cache::finalize_file(plan.dflash);
+                finalize(plan.dflash);
            }
        });
    }
@@ -661,9 +669,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan.eagle3, opts, [&]() {
            // only fall back to the discovered Eagle3 sidecar when no draft was explicitly provided
            if (params.speculative.draft.mparams.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.eagle3);
+                params.speculative.draft.mparams.path = finalize(plan.eagle3);
            } else {
-                hf_cache::finalize_file(plan.eagle3);
+                finalize(plan.eagle3);
            }
        });
    }
@@ -671,9 +679,9 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan.dspark, opts, [&]() {
            // only fall back to the discovered DSpark sidecar when no draft was explicitly provided
            if (params.speculative.draft.mparams.empty()) {
-                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.dspark);
+                params.speculative.draft.mparams.path = finalize(plan.dspark);
            } else {
-                hf_cache::finalize_file(plan.dspark);
+                finalize(plan.dspark);
            }
        });
    }
@@ -681,7 +689,7 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        tasks.emplace_back(plan.preset, opts, [&]() {
            // if HF repo is a preset repo, we simply run server in router mode with the preset.ini file
            params.models_preset_hf = params.model.hf_repo; // only for showing a warning
-            params.models_preset    = hf_cache::finalize_file(plan.preset);
+            params.models_preset    = finalize(plan.preset);
            // clear the model so the server starts in router mode
            params.model.path.clear();
            params.model.hf_repo.clear();
@@ -428,6 +428,41 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
            if (message.contains("tool_call_id")) {
                msg.tool_call_id = message.at("tool_call_id");
            }
+            {
+                json extra = json::object();
+                for (const auto & [k, v] : message.items()) {
+                    if (k != "role" && k != "content" && k != "tool_calls" && k != "reasoning_content" &&
+                        k != "name" && k != "tool_call_id") {
+                        extra[k] = v;
+                    }
+                }
+                if (has_tool_calls) {
+                    // per tool call: keys besides type/function/id, and function keys besides name/arguments
+                    json tc_extra = json::array();
+                    bool any      = false;
+                    for (const auto & tool_call : message.at("tool_calls")) {
+                        json e = json::object();
+                        for (const auto & [k, v] : tool_call.items()) {
+                            if (k != "type" && k != "function" && k != "id") {
+                                e[k] = v;
+                            }
+                        }
+                        for (const auto & [k, v] : tool_call.at("function").items()) {
+                            if (k != "name" && k != "arguments") {
+                                e["function"][k] = v;
+                            }
+                        }
+                        any = any || !e.empty();
+                        tc_extra.push_back(e);
+                    }
+                    if (any) {
+                        extra["__tool_call_extras"] = tc_extra;
+                    }
+                }
+                if (!extra.empty()) {
+                    msg.extra_json = extra.dump();
+                }
+            }

            msgs.push_back(msg);
        }
@@ -516,6 +551,50 @@ static json render_message_to_json(const std::vector<common_chat_msg> & msgs, co
    return messages_inp_normalizer(c).normalize(messages);
 }

+// DeepSeek V4.1 reads non-OpenAI message keys and joins content blocks with "\n\n" around images,
+// so keep media markers as separate text parts and restore the extra keys
+static json render_message_to_json_deepseek_v4_1(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
+    json messages = json::array();
+    for (const auto & msg : msgs) {
+        json jmsg;
+        if (msg.contains_media() && msg.content.empty()) {
+            common_chat_msg typed = msg;
+            for (auto & part : typed.content_parts) {
+                part.type = "text";
+            }
+            jmsg = typed.to_json_oaicompat(/* concat_typed_text= */ false);
+        } else {
+            jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
+        }
+        if (!msg.extra_json.empty()) {
+            const json extra = json::parse(msg.extra_json);
+            for (const auto & [k, v] : extra.items()) {
+                if (k == "__tool_call_extras") {
+                    if (!jmsg.contains("tool_calls")) {
+                        continue;
+                    }
+                    auto & tcs = jmsg.at("tool_calls");
+                    for (size_t i = 0; i < tcs.size() && i < v.size(); i++) {
+                        for (const auto & [tk, tv] : v[i].items()) {
+                            if (tk == "function") {
+                                for (const auto & [fk, fv] : tv.items()) {
+                                    tcs[i]["function"][fk] = fv;
+                                }
+                            } else {
+                                tcs[i][tk] = tv;
+                            }
+                        }
+                    }
+                } else if (!jmsg.contains(k)) {
+                    jmsg[k] = v;
+                }
+            }
+        }
+        messages.push_back(std::move(jmsg));
+    }
+    return messages_inp_normalizer(c).normalize(messages);
+}
+
 // DEPRECATED: only used in tests
 json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
    jinja::caps c;
@@ -1208,11 +1287,13 @@ static common_chat_params_init_fn common_chat_template_detect_params_init(const
    // DeepSeek V3.2/V4 format detection: template defines dsml_token and uses it for tool calls.
    // The template source contains the token as a variable assignment, not as a literal in markup.
    // V3.2 names the tool call block "function_calls", V4 names it "tool_calls".
-    if (src.find("dsml_token") != std::string::npos &&
-        src.find("DSML") != std::string::npos &&
-        (src.find("function_calls") != std::string::npos ||
-         src.find("tool_calls") != std::string::npos)) {
-        LOG_DBG("Using specialized template: DeepSeek V3.2/V4\n");
+    // V4.1 tag names have a leading space and its template does not name the dsml_token variable.
+    if ((src.find("dsml_token") != std::string::npos &&
+         src.find("DSML") != std::string::npos &&
+         (src.find("function_calls") != std::string::npos ||
+          src.find("tool_calls") != std::string::npos)) ||
+        src.find("｜DSML｜ calls>") != std::string::npos) {
+        LOG_DBG("Using specialized template: DeepSeek V3.2/V4/V4.1\n");
        return common_chat_params_init_deepseek_v3_2;
    }

@@ -1313,7 +1394,9 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
        workaround::trim_all_content(trimmed_messages);
        messages_to_render = &trimmed_messages;
    }
-    params.messages              = render_message_to_json(*messages_to_render, tmpl.original_caps());
+    params.messages              = src.find("｜DSML｜ calls>") != std::string::npos ?
+                                       render_message_to_json_deepseek_v4_1(*messages_to_render, tmpl.original_caps()) :
+                                       render_message_to_json(*messages_to_render, tmpl.original_caps());
    params.tool_choice           = inputs.tool_choice;
    params.reasoning_format      = inputs.reasoning_format;
    params.enable_thinking       = inputs.enable_thinking;
@@ -94,6 +94,8 @@ struct common_chat_msg {
    std::string                               reasoning_content;
    std::string                               tool_name;
    std::string                               tool_call_id;
+    // non-OpenAI message keys as a JSON object string, only for templates that read them (DeepSeek V4.1)
+    std::string                               extra_json;

    common_json to_json_oaicompat(bool concat_typed_text = false) const;

@@ -26,6 +26,7 @@
 #include <fstream>
 #include <iostream>
 #include <iterator>
+#include <random>
 #include <regex>
 #include <sstream>
 #include <string>
@@ -741,50 +742,56 @@ static inline bool glob_class_match(const char c, const char * pattern, const ch
 }

 // simple glob: * matches non-/ chars, ** matches anything including /, [] matches character class
-static inline bool glob_match(const char * pattern, const char * str) {
-    if (*pattern == '\0') {
-        return *str == '\0';
-    }
-    if (pattern[0] == '*' && pattern[1] == '*') {
-        const char * p = pattern + 2;
-        if (glob_match(p, str)) return true;
-        if (*str != '\0') return glob_match(pattern, str + 1);
+// only wildcards recurse, too many wildcards: no match
+static bool glob_match(const char * pattern, const char * str, int depth = 0) {
+    if (depth >= COMMON_RECURSIVE_LIMIT) {
        return false;
    }
-    if (*pattern == '*') {
-        const char * p = pattern + 1;
-        for (; *str != '\0' && *str != '/'; str++) {
-            if (glob_match(p, str)) return true;
-        }
-        return glob_match(p, str);
-    }
-    if (*pattern == '?' && *str != '\0' && *str != '/') {
-        return glob_match(pattern + 1, str + 1);
-    }
-    if (*pattern == '[') {
-        const char * class_end = pattern + 1;
-        // If first character after '[' is ']' or '-', treat it as literal
-        if (*class_end == ']' || *class_end == '-') {
-            class_end++;
-        }
-        while (*class_end != '\0' && *class_end != ']') {
-            class_end++;
-        }
-        if (*class_end == ']') {
-            if (*str == '\0') return false;
-            bool matched = glob_class_match(*str, pattern + 1, class_end);
-            return matched && glob_match(class_end + 1, str + 1);
-        } else {
-            if (*str == '[') {
-                return glob_match(pattern + 1, str + 1);
+    while (*pattern != '\0') {
+        if (pattern[0] == '*' && pattern[1] == '*') {
+            const char * p = pattern + 2;
+            for (;; str++) {
+                if (glob_match(p, str, depth + 1)) return true;
+                if (*str == '\0') return false;
            }
-            return false;
        }
+        if (*pattern == '*') {
+            const char * p = pattern + 1;
+            for (; *str != '\0' && *str != '/'; str++) {
+                if (glob_match(p, str, depth + 1)) return true;
+            }
+            return glob_match(p, str, depth + 1);
+        }
+        if (*pattern == '?' && *str != '\0' && *str != '/') {
+            pattern++;
+            str++;
+            continue;
+        }
+        if (*pattern == '[') {
+            const char * class_end = pattern + 1;
+            // If first character after '[' is ']' or '-', treat it as literal
+            if (*class_end == ']' || *class_end == '-') {
+                class_end++;
+            }
+            while (*class_end != '\0' && *class_end != ']') {
+                class_end++;
+            }
+            if (*class_end == ']') {
+                if (*str == '\0' || !glob_class_match(*str, pattern + 1, class_end)) return false;
+                pattern = class_end + 1;
+                str++;
+                continue;
+            }
+            if (*str != '[') return false;
+            pattern++;
+            str++;
+            continue;
+        }
+        if (*pattern != *str) return false;
+        pattern++;
+        str++;
    }
-    if (*pattern == *str) {
-        return glob_match(pattern + 1, str + 1);
-    }
-    return false;
+    return *str == '\0';
 }

 bool glob_match(const std::string & pattern, const std::string & str) {
@@ -917,9 +924,13 @@ std::string fs_path_to_utf8(const std::filesystem::path & path) {
 }

 void fs_write_atomic(const std::filesystem::path & path, const std::string & data) {
+    // a random suffix gives every writer its own temporary file, across processes and containers
+    std::random_device rd;
+    const uint64_t id = ((uint64_t) rd() << 32) | rd();
+
    std::error_code ec;
    std::filesystem::path path_tmp = path;
-    path_tmp += ".tmp";
+    path_tmp += "." + std::to_string(id) + ".tmp";

    if (path.has_parent_path()) {
        std::filesystem::create_directories(path.parent_path(), ec);
@@ -1175,6 +1186,7 @@ static const std::map<common_decision_type, std::string> COMMON_DECISION_TYPE_NA
    { COMMON_DECISION_TYPE_LFM2_D1,        "lfm2-d1"       },
    { COMMON_DECISION_TYPE_LFM2_D1_OMNI,   "lfm2-d1-omni"  },
    { COMMON_DECISION_TYPE_GRANITE_SWITCH, "granite-switch"},
+    { COMMON_DECISION_TYPE_H2O_LIGHTNING,  "h2o-lightning" },
    // for future models, ALWAYS add org name prefix to avoid name collisions, e.g. "orgname-modelname"
 };

@@ -1234,6 +1246,9 @@ common_gguf_info common_get_gguf_info(const std::string & fname) {
        info.n_ctx_train = gguf_get_val_u32(gguf_ctx.get(), ctx_id);
    }

+    // same key as llama_model_chat_template(model, nullptr)
+    info.has_chat_tmpl = gguf_find_key(gguf_ctx.get(), "tokenizer.chat_template") >= 0;
+
    return info;
 }

@@ -31,6 +31,8 @@
 #define DIRECTORY_SEPARATOR '/'
 #endif // _WIN32

+#define COMMON_RECURSIVE_LIMIT 128
+
 #define COM_DBG(fmt, ...) LOG_DBG("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
 #define COM_TRC(fmt, ...) LOG_TRC("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
 #define COM_INF(fmt, ...) LOG_INF("cmn  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
@@ -975,6 +977,7 @@ enum common_decision_type {
    COMMON_DECISION_TYPE_LFM2_D1, // same as openjev, the labels depend on the question type
    COMMON_DECISION_TYPE_LFM2_D1_OMNI, // same as laya, other prompt layout
    COMMON_DECISION_TYPE_GRANITE_SWITCH, // same as openjev, run by the decision aLoRA of granite switch
+    COMMON_DECISION_TYPE_H2O_LIGHTNING, // same as openjev, the labels are read after an answer prefill
    COMMON_DECISION_TYPE_UNKNOWN, // a decision model of a type that is not supported
 };

@@ -984,6 +987,7 @@ common_decision_type common_get_decision_type(const struct llama_model * model);
 struct common_gguf_info {
    common_decision_type decision_type = COMMON_DECISION_TYPE_UNKNOWN; // UNKNOWN if the file is missing, unreadable, or invalid
    uint32_t             n_ctx_train   = 0;                            // 0 if unknown
+    bool                 has_chat_tmpl = false;                        // has the default tokenizer.chat_template
 };

 common_gguf_info common_get_gguf_info(const std::string & fname);
@@ -4,6 +4,7 @@
 #include "common.h"
 #include "log.h"
 #include "download.h"
+#include "file-lock.h"
 #include "hf-cache.h"
 #include "json.h"

@@ -77,6 +78,18 @@ std::pair<std::string, std::string> common_download_split_repo_tag(const std::st
    return {hf_repo, tag};
 }

+// file name from url, without the query string
+static std::string url_get_filename(const std::string & url) {
+    std::string filename = url;
+    if (auto pos = filename.rfind('/'); pos != std::string::npos) {
+        filename = filename.substr(pos + 1);
+    }
+    if (auto pos = filename.find('?'); pos != std::string::npos) {
+        filename = filename.substr(0, pos);
+    }
+    return filename;
+}
+
 class ProgressBar : public common_download_callback {
    static inline std::mutex mutex;
    static inline std::map<const ProgressBar *, int> lines;
@@ -100,14 +113,8 @@ public:
    ProgressBar() = default;

    void on_start(const common_download_progress & p) override {
-        filename = p.url;
+        filename = url_get_filename(p.url);

-        if (auto pos = filename.rfind('/'); pos != std::string::npos) {
-            filename = filename.substr(pos + 1);
-        }
-        if (auto pos = filename.find('?'); pos != std::string::npos) {
-            filename = filename.substr(0, pos);
-        }
        for (size_t i = 0; i < filename.size(); ++i) {
            if ((filename[i] & 0xC0) != 0x80) {
                if (len++ == 39) {
@@ -257,15 +264,62 @@ static bool common_pull_file(httplib::Client & cli,
 static int common_download_file_single_online(const std::string & url,
                                              const std::string & path,
                                              const common_download_opts & opts,
-                                              bool skip_etag) {
+                                              bool skip_etag,
+                                              const std::string & lock_path,
+                                              const std::string & alt_path) {
    static const int max_attempts        = 3;
    static const int retry_delay_seconds = 2;

-    const bool file_exists = std::filesystem::exists(std::filesystem::u8path(path));
+    const std::string path_temporary = path + PARTIAL_SUFFIX;
+    const std::string path_lock = lock_path.empty() ? path + ".lock" : lock_path;
+
+    common_download_progress p;
+    p.url = url;
+
+    if (opts.callback) {
+        opts.callback->on_start(p);
+    }
+
+    // the completion callback may itself take the same lock (the HF cache
+    // finalizes blobs under it), so every exit goes through finish()
+    common_file_lock lock(path_lock);
+    auto finish = [&](int status) {
+        lock.close();
+        if (opts.callback) {
+            opts.callback->on_done(p, is_http_status_ok(status));
+        }
+        return status;
+    };
+
+    // alt_path is the HF snapshot: the blob may have been moved there when symlinks are unavailable
+    auto cache_hit = [&] {
+        return std::filesystem::exists(std::filesystem::u8path(path)) ||
+               (!alt_path.empty() && std::filesystem::exists(std::filesystem::u8path(alt_path)));
+    };
+
+    if (skip_etag && cache_hit()) {
+        LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
+        return finish(304); // 304 Not Modified - fake cached response
+    }
+
+    // take the lock before touching the cache, so two processes cannot race on the same file
+    bool waited = false;
+    if (!lock.acquire([&] {
+            if (!waited) {
+                LOG_INF("%s: file '%s' is being downloaded by another process, waiting...\n", __func__, url_get_filename(url).c_str());
+                waited = true;
+            }
+            return !(opts.callback && opts.callback->is_cancelled());
+        })) {
+        return finish(-1);
+    }
+
+    // under the lock: another process may have finished while we waited
+    bool file_exists = cache_hit();

    if (file_exists && skip_etag) {
        LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
-        return 304; // 304 Not Modified - fake cached response
+        return finish(304); // 304 Not Modified - fake cached response
    }

    auto [cli, parts] = common_http_client(url);
@@ -294,9 +348,9 @@ static int common_download_file_single_online(const std::string & url,
        LOG_TRC("%s: HEAD failed, status: %d\n", __func__, head ? head->status : -1);
        if (file_exists) {
            LOG_TRC("%s: using cached file (HEAD failed): %s\n", __func__, path.c_str());
-            return 304; // 304 Not Modified - fake cached response
+            return finish(304); // 304 Not Modified - fake cached response
        }
-        return head ? head->status : -1;
+        return finish(head ? head->status : -1);
    }

    std::string etag;
@@ -304,8 +358,6 @@ static int common_download_file_single_online(const std::string & url,
        etag = head->get_header_value("ETag");
    }

-    common_download_progress p;
-    p.url = url;
    if (head->has_header("Content-Length")) {
        try {
            p.total = std::stoull(head->get_header_value("Content-Length"));
@@ -322,16 +374,16 @@ static int common_download_file_single_online(const std::string & url,
    if (file_exists) {
        if (etag.empty()) {
            LOG_DBG("%s: using cached file (no server etag): %s\n", __func__, path.c_str());
-            return 304; // 304 Not Modified - fake cached response
+            return finish(304); // 304 Not Modified - fake cached response
        }
        if (!last_etag.empty() && last_etag == etag) {
            LOG_DBG("%s: using cached file (same etag): %s\n", __func__, path.c_str());
-            return 304; // 304 Not Modified - fake cached response
+            return finish(304); // 304 Not Modified - fake cached response
        }
        // pass this point, the file exists but is different from the server version, so we need to redownload it
        if (remove(path.c_str()) != 0) {
            LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-            return -1;
+            return finish(-1);
        }
    }

@@ -341,7 +393,6 @@ static int common_download_file_single_online(const std::string & url,
    }

    bool success = false;
-    const std::string path_temporary = path + PARTIAL_SUFFIX;
    int delay = retry_delay_seconds;

    if (!skip_etag && !etag.empty()) {
@@ -349,7 +400,7 @@ static int common_download_file_single_online(const std::string & url,
            LOG_WRN("%s: partial file does not match the server etag, discarding it: %s\n", __func__, path_temporary.c_str());
            if (remove(path_temporary.c_str()) != 0) {
                LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
-                return -1;
+                return finish(-1);
            }
        }
        write_etag(path_temporary, etag);
@@ -359,7 +410,7 @@ static int common_download_file_single_online(const std::string & url,
        // report the resume offset, so a resumed file does not start at 0
        p.downloaded = supports_ranges && std::filesystem::exists(path_temporary) ?
                       std::filesystem::file_size(path_temporary) : 0;
-        opts.callback->on_start(p);
+        opts.callback->on_update(p);
    }

    for (int i = 0; i < max_attempts; ++i) {
@@ -403,16 +454,12 @@ static int common_download_file_single_online(const std::string & url,
        }
    }

-    if (opts.callback) {
-        opts.callback->on_done(p, success);
-    }
    // a stopped download keeps path_temporary, so the next attempt resumes from it
    if (!success) {
        LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
-        return -1; // max attempts reached
    }

-    return head->status;
+    return finish(success ? head->status : -1);
 }

 std::pair<long, std::vector<char>> common_remote_get_content(const std::string          & url,
@@ -452,14 +499,16 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string
 int common_download_file_single(const std::string & url,
                                const std::string & path,
                                const common_download_opts & opts,
-                                bool skip_etag) {
+                                bool skip_etag,
+                                const std::string & lock_path,
+                                const std::string & alt_path) {
    if (!opts.offline) {
        ProgressBar tty_cb;
        common_download_opts online_opts = opts;
        if (!online_opts.callback) {
            online_opts.callback = &tty_cb;
        }
-        return common_download_file_single_online(url, path, online_opts, skip_etag);
+        return common_download_file_single_online(url, path, online_opts, skip_etag, lock_path, alt_path);
    }

    if (!std::filesystem::exists(std::filesystem::u8path(path))) {
@@ -821,7 +870,7 @@ void common_download_run_tasks(const std::vector<common_download_task> & tasks)
    for (const auto & task : tasks) {
        futures.push_back(std::async(std::launch::async,
            [&task]() {
-                return common_download_file_single(task.url, task.local_path, task.opts, task.is_hf);
+                return common_download_file_single(task.url, task.local_path, task.opts, task.is_hf, task.lock_path, task.alt_path);
            }
        ));
    }
@@ -69,12 +69,15 @@ struct common_download_task {
    std::string local_path;
    std::function<void()> on_done;
    bool is_hf = false;
+    std::string lock_path;
+    std::string alt_path; // HF snapshot path, also treated as cached (the blob may be moved there)

    common_download_task() = default;
    common_download_task(hf_cache::hf_file f,
            const common_download_opts & opts,
            std::function<void()> on_done = nullptr)
-        : opts(opts), url(f.url), local_path(f.local_path), on_done(on_done), is_hf(true) {}
+        : opts(opts), url(f.url), local_path(f.local_path), on_done(on_done), is_hf(true),
+          lock_path(hf_cache::get_lock_path(f.repo_id, f.oid)), alt_path(f.final_path) {}
 };

 void common_download_run_tasks(const std::vector<common_download_task> & tasks);
@@ -92,10 +95,14 @@ std::string common_download_resolve_path(const std::string & hf_repo_with_tag, c
 // download single file from url to local path
 // returns status code or -1 on error
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
+// lock_path: lock file used to serialize downloads of the same file; when empty, "<path>.lock" is used
+// alt_path: additional path treated as present in cache (see cache_hit in download.cpp)
 int common_download_file_single(const std::string & url,
                                const std::string & path,
                                const common_download_opts & opts = {},
-                                bool skip_etag = false);
+                                bool skip_etag = false,
+                                const std::string & lock_path = "",
+                                const std::string & alt_path = "");

 // resolve and download model from Docker registry
 // return local path to downloaded model file
@@ -0,0 +1,179 @@
+#include "file-lock.h"
+
+#include "log.h"
+
+#include <chrono>
+#include <filesystem>
+#include <thread>
+
+#include <fcntl.h>
+#include <sys/stat.h>
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#include <io.h>
+#else
+#include <sys/file.h>
+#include <unistd.h>
+#include <cerrno>
+#include <cstring>
+#endif
+
+common_file_lock::common_file_lock(const std::string & lock_path) {
+    this->lock_path = lock_path;
+}
+
+#if defined(_WIN32)
+// the \\?\ prefix (for paths longer than 255 chars) is only accepted on absolute paths
+// with backslash separators: normalize first, then check the length of the absolute form
+static std::wstring as_extended_path(const std::string & path) {
+    // convert from the narrow encoding std::filesystem uses on this system
+    int wlen = MultiByteToWideChar(CP_ACP, 0, path.c_str(), (int) path.size(), nullptr, 0);
+    if (wlen <= 0) {
+        return std::wstring(path.begin(), path.end()); // best effort for exotic encodings
+    }
+    std::wstring wide(wlen, L'\0');
+    MultiByteToWideChar(CP_ACP, 0, path.c_str(), (int) path.size(), wide.data(), wlen);
+
+    if (wide.rfind(L"\\\\?\\", 0) == 0) {
+        return wide;
+    }
+
+    // GetFullPathNameW resolves relative paths and normalizes separators
+    DWORD size = GetFullPathNameW(wide.c_str(), 0, nullptr, nullptr);
+    if (size == 0) {
+        return wide;
+    }
+    std::wstring full(size, L'\0');
+    DWORD written = GetFullPathNameW(wide.c_str(), size, full.data(), nullptr);
+    if (written == 0 || written >= size) {
+        return wide;
+    }
+    full.resize(written);
+
+    if (full.size() <= 255) {
+        return wide; // short enough: open the path exactly as given
+    }
+    if (full.rfind(L"\\\\", 0) == 0) {
+        return L"\\\\?\\UNC\\" + full.substr(2); // \\server\share -> \\?\UNC\server\share
+    }
+    return L"\\\\?\\" + full;
+}
+#endif
+
+common_file_lock::status common_file_lock::try_lock() {
+    if (lock_path.empty()) {
+        return status::acquired; // no lock path, nothing to lock
+    }
+#if defined(_WIN32)
+    if (fd < 0) {
+        // opened with a wide path: extended paths cannot go through the narrow CRT
+        std::wstring lock_path_w = as_extended_path(lock_path);
+        fd = _wopen(lock_path_w.c_str(), _O_RDWR | _O_CREAT | _O_BINARY | _O_NOINHERIT, _S_IREAD | _S_IWRITE);
+        if (fd < 0) {
+            LOG_ERR("%s: cannot open lock file: %s\n", __func__, lock_path.c_str());
+            return status::error;
+        }
+    }
+    // lock the first byte, the file itself stays empty
+    OVERLAPPED ov = {};
+    locked = LockFileEx((HANDLE) _get_osfhandle(fd),
+                        LOCKFILE_EXCLUSIVE_LOCK | LOCKFILE_FAIL_IMMEDIATELY, 0, 1, 0, &ov);
+    if (locked) {
+        return status::acquired;
+    }
+    if (GetLastError() == ERROR_LOCK_VIOLATION) {
+        return status::busy;
+    }
+    LOG_ERR("%s: file locking failed (%d)\n", __func__, (int) GetLastError());
+    return status::error;
+#else
+    if (fd < 0) {
+        int open_flags = O_RDWR | O_CREAT | O_CLOEXEC;
+#ifdef O_NOFOLLOW
+        open_flags |= O_NOFOLLOW;
+#endif
+        // group-writable, so users sharing a cache can wait on the lock
+        fd = open(lock_path.c_str(), open_flags, 0664);
+        if (fd < 0) {
+            LOG_ERR("%s: cannot open lock file: %s\n", __func__, lock_path.c_str());
+            return status::error;
+        }
+        // open() filters the mode through the umask
+        (void) fchmod(fd, 0664);
+    }
+
+    if (flock(fd, LOCK_EX | LOCK_NB) == 0) {
+        locked = true;
+        return status::acquired;
+    }
+    // EACCES is reported as contention on some filesystems
+    if (errno == EWOULDBLOCK || errno == EAGAIN || errno == EACCES) {
+        return status::busy;
+    }
+    // flock is not implemented (some NFS/FUSE mounts): report it instead of racing
+    LOG_ERR("%s: file locking failed (%s)\n", __func__, strerror(errno));
+    return status::error;
+#endif
+}
+
+bool common_file_lock::acquire(const std::function<bool()> & keep_waiting) {
+    // the lock file is opened by try_lock, its parent must exist by then
+    if (!lock_path.empty()) {
+        std::error_code ec;
+        auto lock_dir = std::filesystem::path(lock_path).parent_path();
+        if (!lock_dir.empty()) {
+            std::filesystem::create_directories(lock_dir, ec);
+            if (ec) {
+                LOG_ERR("%s: unable to create lock directory '%s': %s\n", __func__, lock_dir.string().c_str(), ec.message().c_str());
+                return false;
+            }
+        }
+    }
+
+    while (true) {
+        switch (try_lock()) {
+            case status::acquired:
+                return true;
+            case status::busy:
+                if (keep_waiting && !keep_waiting()) {
+                    return false;
+                }
+                std::this_thread::sleep_for(std::chrono::milliseconds(100));
+                continue;
+            case status::error:
+                return false;
+        }
+    }
+}
+
+// the lock file is never deleted: another process could then create a fresh
+// file at the same path and lock that
+void common_file_lock::close() {
+#if defined(_WIN32)
+    if (fd >= 0) {
+        if (locked) {
+            OVERLAPPED ov = {};
+            UnlockFileEx((HANDLE) _get_osfhandle(fd), 0, 1, 0, &ov);
+        }
+        _close(fd);
+    }
+#else
+    if (fd >= 0) {
+        if (locked) {
+            flock(fd, LOCK_UN);
+        }
+        ::close(fd);
+    }
+#endif
+    fd     = -1;
+    locked = false;
+}
+
+common_file_lock::~common_file_lock() {
+    close();
+}
@@ -0,0 +1,36 @@
+#pragma once
+
+#include <functional>
+#include <string>
+
+// Cross-process exclusive lock (flock on POSIX, LockFileEx on Windows).
+// Lazy: acquire() opens the lock file, so a read-only cache works if the lock is never acquired.
+class common_file_lock {
+public:
+    common_file_lock(const std::string & lock_path);
+    ~common_file_lock();
+
+    // take the lock, waiting while another process holds it;
+    // keep_waiting is polled before each retry, false cancels the wait.
+    // Also returns false when locking is unavailable.
+    bool acquire(const std::function<bool()> & keep_waiting = {});
+
+    void close();
+
+    common_file_lock(const common_file_lock &) = delete;
+    common_file_lock & operator=(const common_file_lock &) = delete;
+
+private:
+    enum class status {
+        acquired,
+        busy,  // another process holds the lock
+        error, // file locking is unavailable: no exclusion is possible
+    };
+
+    status try_lock();
+
+    std::string lock_path;
+
+    int  fd     = -1;
+    bool locked = false;
+};
@@ -2,6 +2,7 @@

 #include "build-info.h"
 #include "common.h"
+#include "file-lock.h"
 #include "log.h"
 #include "http.h"
 #include "json.h"
@@ -161,6 +162,14 @@ static bool is_valid_oid(const std::string & oid) {
    return is_hex_string(oid, 40) || is_hex_string(oid, 64);
 }

+std::string get_lock_path(const std::string & repo_id, const std::string & oid) {
+    if (!is_valid_repo_id(repo_id) || !is_valid_oid(oid)) {
+        return {};
+    }
+    fs::path lock_path = fs::path(get_cache_path()) / ".locks" / repo_to_folder_name(repo_id) / (oid + ".lock");
+    return lock_path.string();
+}
+
 static bool is_valid_subpath(const fs::path & path, const fs::path & subpath) {
    if (subpath.is_absolute()) {
        return false; // never do a / b with b absolute
@@ -437,7 +446,7 @@ hf_files get_cached_files(const std::string & repo_id) {
    return files;
 }

-std::string finalize_file(const hf_file & file) {
+std::string finalize_file(const hf_file & file, const std::function<bool()> & keep_waiting) {
    static std::atomic<bool> symlinks_disabled{false};

    std::error_code ec;
@@ -452,6 +461,24 @@ std::string finalize_file(const hf_file & file) {
        return file.final_path;
    }

+    // the lock file is shared with huggingface_hub: another process may be
+    // finalizing the same blob
+    common_file_lock lock(get_lock_path(file.repo_id, file.oid));
+    if (!lock.acquire(keep_waiting)) {
+        // without the lock, another process could move the blob while we check it
+        throw std::runtime_error("failed to lock cache for finalization: " + file.final_path);
+    }
+
+    // acquire() only polls keep_waiting while waiting for a busy lock
+    if (keep_waiting && !keep_waiting()) {
+        throw std::runtime_error("download cancelled before finalization: " + file.final_path);
+    }
+
+    // re-check under the lock: the winner may have created the link already
+    if (fs::exists(final_path, ec)) {
+        return file.final_path;
+    }
+
    fs::create_directories(final_path.parent_path(), ec);

    if (!symlinks_disabled) {
@@ -1,5 +1,6 @@
 #pragma once

+#include <functional>
 #include <string>
 #include <vector>

@@ -26,8 +27,14 @@ hf_files get_repo_files(

 hf_files get_cached_files(const std::string & repo_id = {});

-// Create snapshot path (link or move/copy) and return it
-std::string finalize_file(const hf_file & file);
+// Lock file path for a blob, shared with the hf CLI: <cache>/.locks/<repo folder>/<oid>.lock
+// Pure path computation: creates nothing. Returns an empty string on invalid repo or oid.
+std::string get_lock_path(const std::string & repo_id, const std::string & oid);
+
+// Create snapshot path (link or move/copy) and return it.
+// Throws when the lock cannot be taken or keep_waiting returns false.
+// A completed snapshot (final_path present) is returned without the lock, so a read-only cache works.
+std::string finalize_file(const hf_file & file, const std::function<bool()> & keep_waiting = {});

 // Remove the entire cached directory for a repo, returns true if removed
 bool remove_cached_repo(const std::string & repo_id);
@@ -1,7 +1,9 @@
 #include "lexer.h"
 #include "runtime.h"
+#include "unicode.h"

 #include <cctype>
+#include <charconv>
 #include <functional>
 #include <map>
 #include <string>
@@ -75,6 +77,24 @@ lexer_result lexer::tokenize(const std::string & source) {
                }
                // add escaped char
                char escaped_char = src[pos++];
+                if (escaped_char == 'x') {
+                    if (pos + 1 >= src.size()) {
+                        throw lexer_exception("incomplete hexadecimal escape", source, pos);
+                    }
+
+                    unsigned int value = 0;
+                    const char * first = &src[pos];
+                    const char * last  = first + 2;
+
+                    auto [ptr, ec] = std::from_chars(first, last, value, 16);
+                    if (ec != std::errc{} || ptr != last) {
+                        throw lexer_exception("invalid hexadecimal escape", source, pos);
+                    }
+
+                    str += common_unicode_cpt_to_utf8(value);
+                    pos += 2;
+                    continue;
+                }
                if (escape_chars.find(escaped_char) == escape_chars.end()) {
                    throw lexer_exception(std::string("unknown escape character \\") + escaped_char, source, pos);
                }
@@ -21,6 +21,17 @@ static bool is_type(const statement_ptr & ptr) {
 class parser {
    const std::vector<token> & tokens;
    size_t current = 0;
+    int depth = 0;
+
+    struct depth_guard {
+        parser & p;
+        depth_guard(parser & p) : p(p) {
+            if (++p.depth > JINJA_RECURSIVE_LIMIT) {
+                throw parser_exception("Parser Error: template is nested more than " + std::to_string(JINJA_RECURSIVE_LIMIT) + " levels deep", p.source, p.peek().pos);
+            }
+        }
+        ~depth_guard() { p.depth--; }
+    };

    std::string source; // for error reporting

@@ -94,6 +105,7 @@ private:
    }

    statement_ptr parse_any() {
+        depth_guard guard(*this);
        size_t start_pos = current;
        switch (peek().t) {
            case token::comment:
@@ -254,6 +266,7 @@ private:
            size_t pos0 = current;
            ++current; // consume {%
            ++current; // consume 'elif'
+            depth_guard guard(*this);
            alternate.push_back(parse_if_statement(pos0)); // nested If
        } else if (is_statement({"else"})) {
            ++current; // consume {%
@@ -330,6 +343,7 @@ private:
    }

    statement_ptr parse_if_expression() {
+        depth_guard guard(*this);
        auto a = parse_logical_or_expression();
        if (is_identifier("if")) {
            // Ternary expression
@@ -373,6 +387,7 @@ private:
    statement_ptr parse_logical_negation_expression() {
        // Try parse unary operators
        if (is_identifier("not")) {
+            depth_guard guard(*this);
            size_t start_pos = current;
            auto op = next();
            return mk_stmt<unary_expression>(start_pos, op, parse_logical_negation_expression());
@@ -458,6 +473,7 @@ private:

    statement_ptr parse_unary_expression() {
        if (is(token::unary_operator)) {
+            depth_guard guard(*this);
            size_t start_pos = current;
            auto op = next();
            return mk_stmt<unary_expression>(start_pos, op, parse_unary_expression());
@@ -477,9 +493,11 @@ private:
        size_t start_pos = current;
        auto expr = mk_stmt<call_expression>(start_pos, std::move(callee), parse_args());
        auto member = parse_member_expression(std::move(expr)); // foo.x().y
-        return is(token::open_paren)
-            ? parse_call_expression(std::move(member)) // foo.x()()
-            : std::move(member);
+        if (!is(token::open_paren)) {
+            return member;
+        }
+        depth_guard guard(*this);
+        return parse_call_expression(std::move(member)); // foo.x()()
    }

    statements parse_args() {
@@ -3,6 +3,7 @@
 #include "value.h"
 #include "utils.h"

+#include <exception>
 #include <string>
 #include <vector>
 #include <memory>
@@ -52,32 +53,39 @@ static void ensure_key_type_allowed(const value & val) {

 // execute with error handling
 value statement::execute(context & ctx) const {
+    std::exception_ptr pending;
+    std::string message;
    try {
        return execute_impl(ctx);
    } catch (const continue_statement::signal & /* ex */) {
-        throw;
+        pending = std::current_exception();
    } catch (const break_statement::signal & /* ex */) {
-        throw;
+        pending = std::current_exception();
    } catch (const rethrown_exception & /* ex */) {
-        throw;
+        pending = std::current_exception();
    } catch (const not_implemented_exception & /* ex */) {
-        throw;
+        pending = std::current_exception();
    } catch (const std::exception & e) {
        const std::string & source = *ctx.src;
        if (source.empty()) {
            std::ostringstream oss;
            oss << "\nError executing " << type() << " at position " << pos << ": " << e.what();
-            throw rethrown_exception(oss.str());
+            message = oss.str();
        } else {
            std::ostringstream oss;
            oss << "\n------------\n";
            oss << "While executing " << type() << " at " << get_line_col(source, pos) << " in source:\n";
            oss << peak_source(source, pos) << "\n";
            oss << "Error: " << e.what();
-            // throw as another exception to avoid repeated formatting
-            throw rethrown_exception(oss.str());
+            message = oss.str();
        }
    }
+    // throw outside the catch: MSVC keeps the unwound frames on the stack while a catch block runs
+    if (pending) {
+        std::rethrow_exception(pending);
+    }
+    // throw as another exception to avoid repeated formatting
+    throw rethrown_exception(message);
 }

 value identifier::execute_impl(context & ctx) const {
@@ -747,7 +755,11 @@ value macro_statement::execute_impl(context & ctx) const {
    std::string name = cast_stmt<identifier>(this->name)->val;

    const func_handler func = [this, name](const func_args & args) -> value {
+        if (args.ctx.macro_depth >= JINJA_RECURSIVE_LIMIT) {
+            throw std::runtime_error("macro calls are nested more than " + std::to_string(JINJA_RECURSIVE_LIMIT) + " levels deep");
+        }
        context macro_ctx(args.ctx); // new scope for macro execution
+        macro_ctx.macro_depth = args.ctx.macro_depth + 1;

        bind_parameters(name, this->args, args, macro_ctx);

@@ -778,7 +790,11 @@ value call_statement::execute_impl(context & ctx) const {
    context caller_ctx(ctx); // new scope for caller execution

    const func_handler func = [this, caller_ctx = std::move(caller_ctx)](const func_args & args) -> value {
+        if (args.ctx.macro_depth >= JINJA_RECURSIVE_LIMIT) {
+            throw std::runtime_error("macro calls are nested more than " + std::to_string(JINJA_RECURSIVE_LIMIT) + " levels deep");
+        }
        context block_ctx(caller_ctx); // new scope for block execution
+        block_ctx.macro_depth = args.ctx.macro_depth + 1;

        bind_parameters("caller", this->caller_args, args, block_ctx);

@@ -58,6 +58,8 @@ struct context {

    bool is_get_stats = false; // whether to collect stats

+    int macro_depth = 0;
+
    visitor_fn visitor;

    // src is optional, used for error reporting
@@ -78,6 +80,7 @@ struct context {
        src(parent.src),
        current_time(parent.current_time),
        is_get_stats(parent.is_get_stats),
+        macro_depth(parent.macro_depth),
        // inherit variables (for example, when entering a new scope)
        env(mk_val<value_object>(*parent.env)) {}

@@ -710,7 +713,7 @@ struct rethrown_exception : public std::exception {

 //////////////////////

-static void gather_string_parts_recursive(const value & val, value_string & parts) {
+static void gather_string_parts_recursive(const value & val, value_string & parts, int depth = 0) {
    // TODO: probably allow print value_none as "None" string? currently this breaks some templates
    if (is_val<value_string>(val)) {
        const auto & str_val = cast_val<value_string>(val)->val_str;
@@ -718,10 +721,12 @@ static void gather_string_parts_recursive(const value & val, value_string & part
    } else if (is_val<value_int>(val) || is_val<value_float>(val) || is_val<value_bool>(val)) {
        std::string str_val = val->as_string().str();
        parts->val_str.append(str_val);
+    } else if (is_val<value_array>(val) && depth >= JINJA_RECURSIVE_LIMIT) {
+        throw std::runtime_error("value is nested more than " + std::to_string(JINJA_RECURSIVE_LIMIT) + " levels deep");
    } else if (is_val<value_array>(val)) {
        auto items = cast_val<value_array>(val)->as_array();
        for (const auto & item : items) {
-            gather_string_parts_recursive(item, parts);
+            gather_string_parts_recursive(item, parts, depth + 1);
        }
    }
 }
@@ -6,6 +6,8 @@
 #include <cstdint>
 #include <cstring>

+#define JINJA_RECURSIVE_LIMIT 128
+
 namespace jinja {

 static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
@@ -6,6 +6,7 @@
 #include "json.h"

 #include <sstream>
+#include <stdexcept>
 #include <string>
 #include <cctype>
 #include <cfloat>
@@ -299,74 +300,53 @@ static value & get_attribute(const value & val, const value & attr, value & defa
    return default_val;
 }

-template<bool is_reject>
-static value selectattr(const func_args & args) {
-    args.ensure_count(2, 4);
-    args.ensure_vals<value_array, value_string, value_string, value_string>(true, true, false, false);
+template<bool is_reject, bool lookup_attr>
+static value select_or_reject(const func_args & args) {
+    // example: array | select("odd") -> test_is_odd(item)
+    // example: array | selectattr("status", "equalto", "active") -> test_is_equalto(item.status, "active")
+    constexpr size_t test_pos = lookup_attr ? 2 : 1;
+    args.ensure_count(test_pos);
+    args.ensure_val<value_array>(args.get_pos(0));
+    value attr;
+    if constexpr (lookup_attr) {
+        attr = args.get_pos(1);
+        args.ensure_val<value_string>(attr);
+    }
+
+    auto & builtins = global_builtins();
+    auto test_it = builtins.end();
+    if (args.count() > test_pos) {
+        args.ensure_val<value_string>(args.get_pos(test_pos));
+        std::string test_name = args.get_pos(test_pos)->as_string().str();
+        test_it = builtins.find("test_is_" + test_name);
+        if (test_it == builtins.end()) {
+            throw raised_exception(args.func_name + ": unknown test '" + test_name + "'");
+        }
+    }

    auto arr = args.get_pos(0)->as_array();
-    auto attribute = args.get_pos(1);
    auto out = mk_val<value_array>();
    value val_default = mk_val<value_undefined>();

-    if (args.count() == 2) {
-        // example: array | selectattr("active")
-        for (const auto & item : arr) {
-            value attr_val = get_attribute(item, attribute, val_default);
-            bool is_selected = attr_val->as_bool();
-            if constexpr (is_reject) is_selected = !is_selected;
-            if (is_selected) out->push_back(item);
+    for (const auto & item : arr) {
+        value val = item;
+        if constexpr (lookup_attr) {
+            val = get_attribute(item, attr, val_default);
        }
-        return out;
-
-    } else if (args.count() == 3) {
-        // example: array | selectattr("equalto", "text")
-        // translated to: test_is_equalto(item, "text")
-        std::string test_name = args.get_pos(1)->as_string().str();
-        value test_val = args.get_pos(2);
-        auto & builtins = global_builtins();
-        auto it = builtins.find("test_is_" + test_name);
-        if (it == builtins.end()) {
-            throw raised_exception("selectattr: unknown test '" + test_name + "'");
-        }
-        auto test_fn = it->second;
-        for (const auto & item : arr) {
+        bool is_selected;
+        if (test_it == builtins.end()) {
+            is_selected = val->as_bool();
+        } else {
            func_args test_args(args.ctx);
-            test_args.push_back(item); // current object
-            test_args.push_back(test_val); // extra argument
-            value test_result = test_fn(test_args);
-            bool is_selected = test_result->as_bool();
-            if constexpr (is_reject) is_selected = !is_selected;
-            if (is_selected) out->push_back(item);
+            test_args.push_back(val);
+            for (size_t i = test_pos + 1; i < args.count(); i++) {
+                test_args.push_back(args.get_pos(i));
+            }
+            is_selected = test_it->second(test_args)->as_bool();
        }
-        return out;
-
-    } else if (args.count() == 4) {
-        // example: array | selectattr("status", "equalto", "active")
-        // translated to: test_is_equalto(item.status, "active")
-        std::string test_name = args.get_pos(2)->as_string().str();
-        auto extra_arg = args.get_pos(3);
-        auto & builtins = global_builtins();
-        auto it = builtins.find("test_is_" + test_name);
-        if (it == builtins.end()) {
-            throw raised_exception("selectattr: unknown test '" + test_name + "'");
-        }
-        auto test_fn = it->second;
-        for (const auto & item : arr) {
-            value attr_val = get_attribute(item, attribute, val_default);
-            func_args test_args(args.ctx);
-            test_args.push_back(attr_val); // attribute value
-            test_args.push_back(extra_arg); // extra argument
-            value test_result = test_fn(test_args);
-            bool is_selected = test_result->as_bool();
-            if constexpr (is_reject) is_selected = !is_selected;
-            if (is_selected) out->push_back(item);
-        }
-        return out;
-    } else {
-        throw raised_exception("selectattr: invalid number of arguments");
+        if constexpr (is_reject) is_selected = !is_selected;
+        if (is_selected) out->push_back(item);
    }
-
    return out;
 }

@@ -425,7 +405,12 @@ const func_builtins & global_builtins() {
            throw raised_exception("Jinja Exception: " + msg);
        }},
        {"dict", toobject},
-        {"namespace", toobject},
+        {"namespace", [](const func_args & args) -> value {
+            // unlike a dict, a namespace has no methods: ns.items is the attribute, as in jinja2
+            value out = toobject(args);
+            cast_val<value_object>(out)->has_builtins = false;
+            return out;
+        }},
        {"strftime_now", [](const func_args & args) -> value {
            args.ensure_vals<value_string>();
            std::string format = args.get_pos(0)->as_string().str();
@@ -1090,10 +1075,10 @@ const func_builtins & value_array_t::get_builtins() const {
            auto arr = slice(val->as_array(), start, stop, step);
            return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(arr)) : mk_val<value_array>(std::move(arr));
        }},
-        {"selectattr", selectattr<false>},
-        {"select", selectattr<false>},
-        {"rejectattr", selectattr<true>},
-        {"reject", selectattr<true>},
+        {"selectattr", select_or_reject<false, true>},
+        {"select", select_or_reject<false, false>},
+        {"rejectattr", select_or_reject<true, true>},
+        {"reject", select_or_reject<true, false>},
        {"join", [](const func_args & args) -> value {
            args.ensure_count(1, 3);
            if (!is_val<value_array>(args.get_pos(0))) {
@@ -1455,7 +1440,8 @@ const func_builtins & value_undefined_t::get_builtins() const {
 //////////////////////////////////


-static value from_json(const common_json & j, bool mark_input) {
+// the runtime walks values with recursion, a deeper input overflows the stack
+static value from_json(const common_json & j, bool mark_input, int depth = 0) {
    if (j.is_null()) {
        return mk_val<value_none>();
    } else if (j.is_boolean()) {
@@ -1470,16 +1456,18 @@ static value from_json(const common_json & j, bool mark_input) {
            str->mark_input();
        }
        return str;
+    } else if ((j.is_array() || j.is_object()) && depth >= JINJA_RECURSIVE_LIMIT) {
+        throw std::invalid_argument("JSON is nested more than " + std::to_string(JINJA_RECURSIVE_LIMIT) + " levels deep");
    } else if (j.is_array()) {
        auto arr = mk_val<value_array>();
        for (const auto & item : j) {
-            arr->push_back(from_json(item, mark_input));
+            arr->push_back(from_json(item, mark_input, depth + 1));
        }
        return arr;
    } else if (j.is_object()) {
        auto obj = mk_val<value_object>();
        for (auto it = j.begin(); it != j.end(); ++it) {
-            obj->insert(it.key(), from_json(it.value(), mark_input));
+            obj->insert(it.key(), from_json(it.value(), mark_input, depth + 1));
        }
        return obj;
    } else {
@@ -1674,6 +1662,8 @@ static void value_to_json_internal(std::ostringstream & oss, const value & val,
            }
        }
        oss << "\"";
+    } else if ((is_val<value_array>(val) || is_val<value_object>(val)) && curr_lvl >= JINJA_RECURSIVE_LIMIT) {
+        throw std::runtime_error("value is nested more than " + std::to_string(JINJA_RECURSIVE_LIMIT) + " levels deep");
    } else if (is_val<value_array>(val)) {
        const auto & arr = val->as_array();
        oss << "[";
@@ -1739,17 +1729,20 @@ std::string value_to_string_repr(const value & val) {
 }

 // stats utility
-void value_t::stats_t::mark_used(value & val, bool deep) {
+void value_t::stats_t::mark_used(value & val, bool deep, int depth) {
    val->stats.used = true;
+    if (deep && (is_val<value_array>(val) || is_val<value_object>(val)) && depth >= JINJA_RECURSIVE_LIMIT) {
+        throw std::runtime_error("value is nested more than " + std::to_string(JINJA_RECURSIVE_LIMIT) + " levels deep");
+    }
    if (deep) {
        if (is_val<value_array>(val)) {
            for (auto & item : val->val_arr) {
-                mark_used(item, deep);
+                mark_used(item, deep, depth + 1);
            }
        } else if (is_val<value_object>(val)) {
            for (auto & pair : val->val_obj) {
-                mark_used(pair.first, deep);
-                mark_used(pair.second, deep);
+                mark_used(pair.first, deep, depth + 1);
+                mark_used(pair.second, deep, depth + 1);
            }
        }
    }
@@ -119,7 +119,7 @@ struct value_t {
        // ops can be builtin calls or operators: "array_access", "object_access"
        std::set<std::string> ops;
        // utility to recursively mark value and its children as used
-        static void mark_used(value & val, bool deep = false);
+        static void mark_used(value & val, bool deep = false, int depth = 0);
    } stats;

    value_t() = default;
@@ -272,15 +272,12 @@ static bool is_reserved_name(const std::string & name) {
    return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
 }

-static std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
 static std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"\\\\]");
 static std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
 static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };

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

@@ -312,6 +309,22 @@ static std::string format_literal(const std::string & literal) {

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

+// same as std::regex_replace with [^a-zA-Z0-9-]+ -> "-", no std::regex: libstdc++ recurses per char
+std::string gbnf_format_rule_name(const std::string & name) {
+    std::string out;
+    bool in_invalid = false;
+    for (char c : name) {
+        if ((c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z') || (c >= '0' && c <= '9') || c == '-') {
+            out += c;
+            in_invalid = false;
+        } else if (!in_invalid) {
+            out += '-';
+            in_invalid = true;
+        }
+    }
+    return out;
+}
+
 static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
    if (pos + 1 >= pattern.length() || pattern[pos] != '\\') {
        return 0;
@@ -347,6 +360,17 @@ private:
    std::unordered_set<std::string> _refs_being_resolved;
    std::vector<std::string> _errors;
    std::vector<std::string> _warnings;
+    int _depth = 0;
+
+    struct depth_guard {
+        int & depth;
+        depth_guard(int & depth) : depth(depth) {
+            if (++depth > COMMON_RECURSIVE_LIMIT) {
+                throw std::invalid_argument("JSON schema is nested more than " + std::to_string(COMMON_RECURSIVE_LIMIT) + " levels deep");
+            }
+        }
+        ~depth_guard() { depth--; }
+    };

    template <typename T>
    static const T & as(const common_chat_schema & node) {
@@ -354,7 +378,7 @@ private:
    }

    std::string _add_rule(const std::string & name, const std::string & rule) {
-        std::string esc_name = regex_replace(name, INVALID_RULE_CHARS_RE, "-");
+        std::string esc_name = gbnf_format_rule_name(name);
        if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) {
            _rules[esc_name] = rule;
            return esc_name;
@@ -482,7 +506,7 @@ private:
                        }
                    }
                    paren_depth++;
-                    if (paren_depth > MAX_PATTERN_DEPTH) {
+                    if (paren_depth > COMMON_RECURSIVE_LIMIT) {
                        throw unsupported_pattern("pattern nesting too deep");
                    }
                    seq.emplace_back("(" + to_rule(transform()) + ")", false);
@@ -644,7 +668,11 @@ private:
        std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
        std::ostringstream out;
        out << "[\"] ( ";
-        std::function<void(size_t)> visit = [&](size_t idx) {
+        // one level per char of the longest name
+        std::function<void(size_t, int)> visit = [&](size_t idx, int depth) {
+            if (depth >= COMMON_RECURSIVE_LIMIT) {
+                throw std::invalid_argument("JSON schema property name is longer than " + std::to_string(COMMON_RECURSIVE_LIMIT) + " chars");
+            }
            const auto & node = trie.nodes[idx];
            std::string rejects;
            auto first = true;
@@ -659,7 +687,7 @@ private:
                out << "[" << c << "]";
                if (!trie.nodes[child].children.empty()) {
                    out << " (";
-                    visit(child);
+                    visit(child, depth + 1);
                    out << ")";
                } else {
                    out << " " << char_rule << "+";
@@ -669,7 +697,7 @@ private:
                out << " | [^\"" << rejects << "] " << char_rule << "*";
            }
        };
-        visit(0);
+        visit(0, 0);

        out << " )";
        if (trie.nodes[0].pattern < 0) {
@@ -682,8 +710,7 @@ private:
    std::string _resolve_ref(const common_chat_schema_ref & schema) {
        auto it = schema.ref.find('#');
        std::string ref_fragment = it != std::string::npos ? schema.ref.substr(it + 1) : schema.ref;
-        static const std::regex nonalphanumeric_regex(R"([^a-zA-Z0-9-]+)");
-        std::string ref_name = "ref" + std::regex_replace(ref_fragment, nonalphanumeric_regex, "-");
+        std::string ref_name = "ref" + gbnf_format_rule_name(ref_fragment);
        if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(schema.ref) == _refs_being_resolved.end()) {
            if (!schema.target) {
                _errors.push_back("Unresolved $ref " + schema.ref);
@@ -833,10 +860,13 @@ public:
        std::unordered_set<std::string> required;
        std::vector<std::pair<std::string, const common_chat_schema *>> properties;
        std::map<std::string, size_t> enum_values;
-        std::function<void(const common_chat_schema &, bool)> add_component = [&](const common_chat_schema & comp, bool is_required) {
+        std::function<void(const common_chat_schema &, bool, int)> add_component = [&](const common_chat_schema & comp, bool is_required, int depth) {
+            if (depth >= COMMON_RECURSIVE_LIMIT) {
+                throw std::invalid_argument("JSON schema is nested more than " + std::to_string(COMMON_RECURSIVE_LIMIT) + " levels deep");
+            }
            if (comp.kind() == common_chat_schema::KIND_REF) {
                if (const auto * target = as<common_chat_schema_ref>(comp).target) {
-                    add_component(*target, is_required);
+                    add_component(*target, is_required, depth + 1);
                }
            } else if (comp.kind() == common_chat_schema::KIND_OBJECT) {
                for (const auto & prop : as<common_chat_schema_object>(comp).properties) {
@@ -854,10 +884,10 @@ public:
        for (const auto & child : schema.children) {
            if (child->kind() == common_chat_schema::KIND_ANY_OF) {
                for (const auto & alt : as<common_chat_schema_any_of>(*child).children) {
-                    add_component(*alt, false);
+                    add_component(*alt, false, 0);
                }
            } else {
-                add_component(*child, true);
+                add_component(*child, true, 0);
            }
        }
        if (!enum_values.empty()) {
@@ -875,6 +905,8 @@ public:
    }

    std::string visit(const common_chat_schema & schema, const std::string & name) {
+        // the $ref targets are visited too, a $ref chain is not bounded by the builder
+        depth_guard guard(_depth);
        std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name;
        std::string sub_name  = name + (name.empty() ? "" : "-");

@@ -19,5 +19,6 @@ struct common_grammar_options {
 };

 std::string gbnf_format_literal(const std::string & literal);
+std::string gbnf_format_rule_name(const std::string & name);

 std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});
@@ -9,6 +9,18 @@
 #include <utility>
 #include <vector>

+namespace {
+struct depth_guard {
+    int & depth;
+    depth_guard(int & depth) : depth(depth) {
+        if (++depth > COMMON_RECURSIVE_LIMIT) {
+            throw std::invalid_argument("JSON schema is nested more than " + std::to_string(COMMON_RECURSIVE_LIMIT) + " levels deep");
+        }
+    }
+    ~depth_guard() { depth--; }
+};
+} // namespace
+
 class common_chat_schema_builder {
    const common_json &      root_;
    common_chat_schema_document & doc_;
@@ -224,7 +236,14 @@ class common_chat_schema_builder {
        return node;
    }

+    int depth_ = 0;
+
    common_chat_schema_ptr build_node(const common_json & schema, const std::string & path) {
+        depth_guard guard(depth_);
+        return build_node_impl(schema, path);
+    }
+
+    common_chat_schema_ptr build_node_impl(const common_json & schema, const std::string & path) {
        if (!schema.is_object()) {
            fail(path, "schema must be an object");
        }
@@ -365,7 +384,9 @@ static common_chat_schema::value_type json_type(const common_json & value) {
    return common_chat_schema::TYPE_OBJECT;
 }

-static common_chat_schema::type_set value_types_impl(const common_chat_schema & s, std::unordered_set<const common_chat_schema *> & visited) {
+// depth counts the $ref, anyOf and allOf hops, a $ref chain is not bounded by the builder
+static common_chat_schema::type_set value_types_impl(const common_chat_schema & s, std::unordered_set<const common_chat_schema *> & visited, int & depth) {
+    depth_guard guard(depth);
    switch (s.kind()) {
        case common_chat_schema::KIND_ANY:
            return common_chat_schema::type_set::all();
@@ -399,21 +420,21 @@ static common_chat_schema::type_set value_types_impl(const common_chat_schema &
                // a cycle contributes no type, to be safe
                return {};
            }
-            auto types = value_types_impl(*target, visited);
+            auto types = value_types_impl(*target, visited, depth);
            visited.erase(target);
            return types;
        }
        case common_chat_schema::KIND_ANY_OF: {
            common_chat_schema::type_set types;
            for (const auto & child : static_cast<const common_chat_schema_any_of &>(s).children) {
-                types |= value_types_impl(*child, visited);
+                types |= value_types_impl(*child, visited, depth);
            }
            return types;
        }
        case common_chat_schema::KIND_ALL_OF: {
            auto types = common_chat_schema::type_set::all();
            for (const auto & child : static_cast<const common_chat_schema_all_of &>(s).children) {
-                types &= value_types_impl(*child, visited);
+                types &= value_types_impl(*child, visited, depth);
            }
            return types;
        }
@@ -423,10 +444,12 @@ static common_chat_schema::type_set value_types_impl(const common_chat_schema &

 common_chat_schema::type_set common_chat_schema::value_types() const {
    std::unordered_set<const common_chat_schema *> visited;
-    return value_types_impl(*this, visited);
+    int depth = 0;
+    return value_types_impl(*this, visited, depth);
 }

-static bool may_be_string_impl(const common_chat_schema & s, std::unordered_set<const common_chat_schema *> & visited) {
+static bool may_be_string_impl(const common_chat_schema & s, std::unordered_set<const common_chat_schema *> & visited, int & depth) {
+    depth_guard guard(depth);
    switch (s.kind()) {
        case common_chat_schema::KIND_STRING:
            return true;
@@ -445,13 +468,13 @@ static bool may_be_string_impl(const common_chat_schema & s, std::unordered_set<
            if (!target || !visited.insert(target).second) {
                return false;
            }
-            bool result = may_be_string_impl(*target, visited);
+            bool result = may_be_string_impl(*target, visited, depth);
            visited.erase(target);
            return result;
        }
        case common_chat_schema::KIND_ANY_OF:
            for (const auto & child : static_cast<const common_chat_schema_any_of &>(s).children) {
-                if (may_be_string_impl(*child, visited)) {
+                if (may_be_string_impl(*child, visited, depth)) {
                    return true;
                }
            }
@@ -463,7 +486,7 @@ static bool may_be_string_impl(const common_chat_schema & s, std::unordered_set<
                if (child->kind() == common_chat_schema::KIND_ANY) {
                    continue;
                }
-                if (!may_be_string_impl(*child, visited)) {
+                if (!may_be_string_impl(*child, visited, depth)) {
                    return false;
                }
                any_string = true;
@@ -477,7 +500,8 @@ static bool may_be_string_impl(const common_chat_schema & s, std::unordered_set<

 bool common_chat_schema::may_be_string() const {
    std::unordered_set<const common_chat_schema *> visited;
-    return may_be_string_impl(*this, visited);
+    int depth = 0;
+    return may_be_string_impl(*this, visited, depth);
 }

 const char * common_chat_schema::kind_name(node_kind kind) {
@@ -1,4 +1,5 @@
 #include "parsers.h"
+#include "log.h"

 // The DeepSeek V4 reference implementation renders consecutive tool results into a single
 // user block, ordered by the tool call order of the preceding assistant message (matched
@@ -75,13 +76,30 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
    // non-thinking generation prompt ends with a bare </think> instead of an empty
    // <think></think> pair.
    const bool is_v4 = tmpl.source().find("function_calls") == std::string::npos;
+    // V4.1: DSML tag names carry a leading space, e.g. <｜DSML｜ calls>
+    const bool is_v41 = is_v4 && tmpl.source().find("｜DSML｜ calls>") != std::string::npos;

    std::optional<json> adjusted_messages;
    if (is_v4) {
        adjusted_messages = deepseek_v4_sort_tool_results(inputs.messages);
    }

-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    // V4.1 can declare tools on a system message instead of top-level "tools": parse calls to those too.
+    json tools = inputs.tools;
+    if (is_v41 && !(tools.is_array() && !tools.empty()) && inputs.messages.is_array()) {
+        for (const auto & msg : inputs.messages) {
+            if (msg.value("role", "") == "system" && msg.contains("tools") && msg.at("tools").is_array()) {
+                if (!tools.is_array()) {
+                    tools = json::array();
+                }
+                for (const auto & tool : msg.at("tools")) {
+                    tools.push_back(tool);
+                }
+            }
+        }
+    }
+
+    auto has_tools           = tools.is_array() && !tools.empty();
    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
@@ -90,17 +108,39 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
    if (is_v4 && has_response_format) {
        additional_context = json{ { "response_format", inputs.json_schema } };
    }
+    // V4.1 raises on anything but low/high/max or 1-100: map minimal -> low, xhigh -> max, medium -> default
+    if (is_v41 && inputs.extra_context.is_object() && inputs.extra_context.contains("reasoning_effort") &&
+        inputs.extra_context.at("reasoning_effort").is_string()) {
+        const std::string effort = inputs.extra_context.at("reasoning_effort").get<std::string>();
+        json mapped;
+        if (effort == "minimal") {
+            mapped = "low";
+        } else if (effort == "xhigh") {
+            mapped = "max";
+        } else if (effort == "medium") {
+            mapped = nullptr;
+        }
+        if (effort == "minimal" || effort == "xhigh" || effort == "medium") {
+            LOG_DBG("%s: DeepSeek V4.1 reasoning_effort '%s' -> %s\n", __func__, effort.c_str(),
+                    mapped.is_null() ? "unset (template default)" : mapped.get<std::string>().c_str());
+            if (!additional_context.has_value()) {
+                additional_context = json::object();
+            }
+            (*additional_context)["reasoning_effort"] = mapped;
+        }
+    }

    const std::string DSML         = "｜DSML｜";
    const std::string THINK_START  = "<think>";
    const std::string THINK_END    = "</think>";
-    const std::string TC_BLOCK     = is_v4 ? "tool_calls" : "function_calls";
-    const std::string FC_START     = "<" + DSML + TC_BLOCK + ">";
-    const std::string FC_END       = "</" + DSML + TC_BLOCK + ">";
-    const std::string INVOKE_START = "<" + DSML + "invoke";
-    const std::string INVOKE_END   = "</" + DSML + "invoke>";
-    const std::string PARAM_START  = "<" + DSML + "parameter";
-    const std::string PARAM_END    = "</" + DSML + "parameter>";
+    const std::string TAG_SP       = is_v41 ? " " : "";
+    const std::string TC_BLOCK     = is_v41 ? "calls" : is_v4 ? "tool_calls" : "function_calls";
+    const std::string FC_START     = "<" + DSML + TAG_SP + TC_BLOCK + ">";
+    const std::string FC_END       = "</" + DSML + TAG_SP + TC_BLOCK + ">";
+    const std::string INVOKE_START = "<" + DSML + TAG_SP + "invoke";
+    const std::string INVOKE_END   = "</" + DSML + TAG_SP + "invoke>";
+    const std::string PARAM_START  = "<" + DSML + TAG_SP + "parameter";
+    const std::string PARAM_END    = "</" + DSML + TAG_SP + "parameter>";
    const std::string GEN_PROMPT   = "<｜Assistant｜>";
    const std::string TC_SEPARATOR = "\n\n";

@@ -152,7 +192,7 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
        // build tool call section first since we might need it in reasoning
        auto tool_choice = p.choice();
        if (has_tool_calls) {
-            foreach_function(inputs.tools, [&](size_t tool_index, const json & tool) {
+            foreach_function(tools, [&](size_t tool_index, const json & tool) {
                const auto & function = tool.at("function");
                std::string  name     = function.at("name");

@@ -223,7 +263,10 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
        }

        if (extract_reasoning && inputs.enable_thinking) {
-            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+            // V4.1 renders a continuation without reasoning as "<｜Assistant｜></think>" even in thinking mode,
+            // so consume the bare THINK_END too or it leaks into the content
+            reasoning = is_v41 ? p.optional((THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END) | p.literal(THINK_END))
+                               : p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
            reasoning_with_tc = THINK_START +
                p.reasoning(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START, THINK_END })) +
                p.space() + obligatory_tool_calls;
@@ -10,10 +10,12 @@
 #include <initializer_list>
 #include <map>
 #include <memory>
-#include <regex>
 #include <set>
 #include <stdexcept>

+// The parser can run with a lower limit than COMMON_RECURSIVE_LIMIT
+static constexpr int RULE_RECURSIVE_LIMIT = 64;
+
 // Trick to catch missing branches
 template <typename T>
 inline constexpr bool is_always_false_v = false;
@@ -186,13 +188,17 @@ std::string common_peg_ast_node::sanitized_text() const {
 }

 void common_peg_ast_arena::visit(common_peg_ast_id id, const common_peg_ast_visitor & visitor) const {
-    if (id == COMMON_PEG_INVALID_AST_ID) {
-        return;
-    }
-    const auto & node = get(id);
-    visitor(node);
-    for (const auto & child : node.children) {
-        visit(child, visitor);
+    // no recursion: the AST is several nodes deep per level of the input
+    std::vector<common_peg_ast_id> stack = { id };
+    while (!stack.empty()) {
+        const auto cur = stack.back();
+        stack.pop_back();
+        if (cur == COMMON_PEG_INVALID_AST_ID) {
+            continue;
+        }
+        const auto & node = get(cur);
+        visitor(node);
+        stack.insert(stack.end(), node.children.rbegin(), node.children.rend());
    }
 }

@@ -736,8 +742,15 @@ struct parser_executor {
    }

    common_peg_parse_result operator()(const common_peg_rule_parser & p) {
+        // only rules recurse, a recursive rule is entered once per input level
+        if (ctx.rule_depth[p.child] >= RULE_RECURSIVE_LIMIT) {
+            throw common_peg_recursion_error("input is nested more than " + std::to_string(RULE_RECURSIVE_LIMIT) + " levels deep");
+        }
+
        // Parse the child
+        ctx.rule_depth[p.child]++;
        auto result = arena.parse(p.child, ctx, start_pos);
+        ctx.rule_depth[p.child]--;

        if (!result.fail()) {
            std::string_view text;
@@ -819,6 +832,10 @@ common_peg_parse_result common_peg_arena::parse(common_peg_parse_context & ctx,
    if (root_ == COMMON_PEG_INVALID_PARSER_ID) {
        throw std::runtime_error("No root parser set");
    }
+
+    // Initialize rule depth for all parsers
+    ctx.rule_depth.assign(parsers_.size(), 0);
+
    return parse(root_, ctx, start);
 }

@@ -1069,11 +1086,6 @@ common_peg_parser operator|(const std::string & str, const common_peg_parser & p
    return operator|(str.c_str(), p);
 }

-static std::string rule_name(const std::string & name) {
-    static const std::regex invalid_rule_chars_re("[^a-zA-Z0-9-]+");
-    return std::regex_replace(name, invalid_rule_chars_re, "-");
-}
-
 common_peg_parser_builder::common_peg_parser_builder() {}

 common_peg_parser common_peg_parser_builder::sequence(const std::vector<common_peg_parser_id> & parsers) {
@@ -1155,14 +1167,14 @@ common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p,
 }

 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const common_peg_parser & p, bool trigger) {
-    auto clean_name = rule_name(name);
+    auto clean_name = gbnf_format_rule_name(name);
    auto rule_id = arena_.add_parser(common_peg_rule_parser{clean_name, p.id(), trigger});
    arena_.add_rule(clean_name, rule_id);
    return ref(clean_name);
 }

 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const std::function<common_peg_parser()> & builder_fn, bool trigger) {
-    auto clean_name = rule_name(name);
+    auto clean_name = gbnf_format_rule_name(name);
    if (arena_.has_rule(clean_name)) {
        return ref(clean_name);
    }
@@ -6,6 +6,7 @@

 #include <memory>
 #include <set>
+#include <stdexcept>
 #include <unordered_map>
 #include <string>
 #include <string_view>
@@ -182,6 +183,10 @@ inline common_peg_parse_flags operator~(common_peg_parse_flags a) {
    return static_cast<common_peg_parse_flags>(~int(a));
 }

+struct common_peg_recursion_error : std::invalid_argument {
+    using std::invalid_argument::invalid_argument;
+};
+
 struct common_peg_parse_context {
    std::string input;               // [h,   e,  l,  l,  o,  _,  w,  o,  r,  l,  d]
    std::vector<llama_token> tokens; // [id, -1, -1, -1, -1, id, -1, -1, -1, -1, -1]
@@ -189,6 +194,7 @@ struct common_peg_parse_context {
    common_peg_ast_arena ast;

    int parse_depth;
+    std::vector<int> rule_depth; // index: rule child id

    common_peg_parse_context(common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_NONE)
        : flags(flags), parse_depth(0) {}
@@ -16,8 +16,10 @@
 #include <cassert>
 #include <cmath>
 #include <cstring>
+#include <fstream>
 #include <iomanip>
 #include <map>
+#include <random>
 #include <cinttypes>

 #define SPC_DBG(fmt, ...) LOG_DBG("spec %12.*s: " fmt, 12, __func__, __VA_ARGS__)
@@ -1009,6 +1011,10 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    bool    is_mrope       = false;
    int32_t selector_top_k = 0;

+    // DFlash2 probabilistic drafting: one generator per sequence and the seed it was last seeded with (-1: reseed)
+    std::vector<std::mt19937> dflash2_rng;
+    std::vector<int64_t>      dflash2_seed;
+
    // draft-dspark: the draft carries a Markov head and uses an anchor-first block layout
    const bool is_dspark;

@@ -1059,6 +1065,8 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

        selector_top_k = llama_model_dflash_selector_top_k(model_dft);
        is_dflash2     = selector_top_k > 0;
+        dflash2_rng.resize(n_seq);
+        dflash2_seed.assign(n_seq, -1);
        char lilicorr_meta[32] = {};
        lilicorr_top_k = llama_model_meta_val_str(model_dft, "lilicorr.candidate_topk", lilicorr_meta, sizeof(lilicorr_meta)) >= 0
                ? std::atoi(lilicorr_meta) : 0;
@@ -1157,6 +1165,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            return;
        }

+        // reseed per request rather than per round, or two identical requests differ
+        dflash2_seed[seq_id] = -1;
+
        const int32_t N = (int32_t) prompt.size();
        if (N <= 0) {
            return;
@@ -1337,6 +1348,21 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
                const float * lattice = llama_get_embeddings_nextn(ctx_dft);
                GGML_ASSERT(lattice && "DFlash2 selector produced no lattice");

+                // DFlash2 can instead sample each step and hand q to the verifier for rejection sampling
+                const bool sample = is_dflash2 && params.probabilistic && dp.result_q;
+
+                // greedy keeps temperature 1, sampling uses the target's
+                const float temp = sample ? dp.temp : 1.0f;
+
+                if (sample && dflash2_seed[seq_id] != dp.seed) {
+                    dflash2_seed[seq_id] = dp.seed;
+                    // the default seed means random; mixed so it differs from the target's, as in spec_retune
+                    dflash2_rng[seq_id].seed(dp.seed == LLAMA_DEFAULT_SEED ? std::random_device{}() : dp.seed ^ 0x85ebca6bu);
+                }
+
+                std::vector<float>   probs;
+                std::vector<int32_t> sorted_idx;
+
                int32_t predecessor = 0;
                for (int32_t i = 1; i < (is_dflash2 ? n_block_tokens : params.n_max + 1); ++i) {
                    const float * row = lattice + (size_t) (beg + i) * n_embd_dec;
@@ -1345,21 +1371,53 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {

                    predecessor = (int32_t) std::distance(scores,
                            std::max_element(scores, scores + top_k));
-                    if (params.p_min > 0.0f) {
-                        // softmax(scores) at the argmax, i.e. 1 / sum(exp(s_k - s_max))
+                    if (params.p_min > 0.0f || sample) {
+                        // softmax(scores / temp) at the argmax, i.e. 1 / sum(exp((s_k - s_max) / temp))
+                        probs.resize(top_k);
                        float sum = 0.0f;
                        for (int32_t k = 0; k < top_k; ++k) {
-                            sum += std::exp(scores[k] - scores[predecessor]);
+                            probs[k] = std::exp((scores[k] - scores[predecessor]) / temp);
+                            sum += probs[k];
                        }
                        if (1.0f / sum < params.p_min) {
                            break;
                        }
+                        for (int32_t k = 0; k < top_k; ++k) {
+                            probs[k] /= sum;
+                        }
+                    }
+                    if (sample) {
+                        // ggml_top_k leaves candidates unordered, so walk them by token id to keep a seed's draft stable
+                        sorted_idx.resize(top_k);
+                        for (int32_t k = 0; k < top_k; ++k) {
+                            sorted_idx[k] = k;
+                        }
+                        std::sort(sorted_idx.begin(), sorted_idx.end(), [row](int32_t a, int32_t b) { return row[a] < row[b]; });
+
+                        // a round-off overshoot keeps the argmax, which has q > 0
+                        double u = std::uniform_real_distribution<double>(0.0, 1.0)(dflash2_rng[seq_id]);
+                        for (const int32_t k : sorted_idx) {
+                            u -= probs[k];
+                            if (u < 0.0) {
+                                predecessor = k;
+                                break;
+                            }
+                        }
+
+                        auto & q = dp.result_q->emplace_back();
+                        q.reserve(top_k);
+                        for (int32_t k = 0; k < top_k; ++k) {
+                            q.push_back({ (llama_token) row[k], scores[k], probs[k] });
+                        }
                    }
                    result.push_back((llama_token) row[predecessor]);
                }

                if (result.size() < (size_t) params.n_min) {
                    result.clear();
+                    if (sample) {
+                        dp.result_q->clear();
+                    }
                }
                continue;
            }
@@ -2152,9 +2210,6 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {

    uint16_t n_draft;

-    bool save_dynamic;
-    bool save_static;
-
    struct seq_info {
        size_t cache_size = 0; // number of tokens in n-gram cache

@@ -2168,54 +2223,72 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
    common_speculative_impl_ngram_cache(
            const common_params_speculative & params,
            uint32_t n_seq,
-            uint16_t n_draft,
-            const std::string & path_static,
-            const std::string & path_dynamic,
-            bool save_dynamic,
-            bool save_static)
+            uint16_t n_draft)
        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, n_seq, n_draft)
        , params(params.ngram_cache)
        , n_draft(n_draft)
-        , save_dynamic(save_dynamic)
-        , save_static(save_static)
    {
+        const auto & ngram = params.ngram_cache;
+
        SPC_TRC("%s", "adding speculative implementation 'ngram-cache'\n");
        SPC_TRC("- n_draft=%d, cache_static=%s, cache_dynamic=%s\n",
                n_draft,
-                path_static.empty() ? "none" : path_static.c_str(),
-                path_dynamic.empty() ? "none" : path_dynamic.c_str());
+                ngram.lookup_cache_static.empty() ? "none" : ngram.lookup_cache_static.c_str(),
+                ngram.lookup_cache_dynamic.empty() ? "none" : ngram.lookup_cache_dynamic.c_str());

        sinfos.resize(n_seq);

-        if (!path_static.empty()) {
+        if (!ngram.lookup_cache_static.empty()) {
            try {
-                auto ngram_cache_static = common_ngram_cache_load(path_static);
+                auto ngram_cache_static = common_ngram_cache_load(ngram.lookup_cache_static);

                for (auto & sinfo : sinfos) {
                    sinfo.ngram_cache_static = ngram_cache_static;
                }
-            } catch (...) {
-                SPC_ERR("failed to open static lookup cache: %s", path_static.c_str());
+            } catch (const std::ifstream::failure &) {
+                SPC_ERR("failed to open static lookup cache: %s\n", ngram.lookup_cache_static.c_str());
                GGML_ABORT("Couldn't read static lookup cache");
            }
        }

-        if (!path_dynamic.empty()) {
+        if (!ngram.lookup_cache_dynamic.empty()) {
            try {
-                auto ngram_cache_dynamic = common_ngram_cache_load(path_dynamic);
+                auto ngram_cache_dynamic = common_ngram_cache_load(ngram.lookup_cache_dynamic);

                for (auto & sinfo : sinfos) {
                    sinfo.ngram_cache_dynamic = ngram_cache_dynamic;
                }
-            } catch (...) {
-                SPC_ERR("failed to open dynamic lookup cache: %s", path_dynamic.c_str());
-                GGML_ABORT("Couldn't read dynamic lookup cache");
+            } catch (const std::ifstream::failure &) {
+                std::error_code ec;
+                const bool exists = std::filesystem::exists(ngram.lookup_cache_dynamic, ec);
+                if (!exists && !ec) {
+                    // missing file is created on destruction
+                    SPC_WRN("dynamic lookup cache file does not exist: %s, starting with empty cache\n",
+                            ngram.lookup_cache_dynamic.c_str());
+                } else {
+                    SPC_ERR("failed to open dynamic lookup cache: %s\n", ngram.lookup_cache_dynamic.c_str());
+                    GGML_ABORT("Couldn't read dynamic lookup cache");
+                }
            }
        }
    }

-    void begin(llama_seq_id /*seq_id*/, const llama_tokens & /*prompt*/) override {
-        // noop
+    ~common_speculative_impl_ngram_cache() override {
+        if (params.lookup_cache_dynamic.empty() || sinfos.empty()) {
+            return;
+        }
+
+        auto & dynamic_cache = sinfos.front().ngram_cache_dynamic;
+        for (auto & sinfo : sinfos) {
+            common_ngram_cache_merge(dynamic_cache, sinfo.ngram_cache_context);
+        }
+
+        SPC_INF("saving dynamic lookup cache at %s\n", params.lookup_cache_dynamic.c_str());
+        common_ngram_cache_save(dynamic_cache, params.lookup_cache_dynamic);
+    }
+
+    void begin(llama_seq_id seq_id, const llama_tokens & /*prompt*/) override {
+        sinfos[seq_id].cache_size = 0;
    }

    void draft_one(
@@ -2226,22 +2299,6 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {

        const auto & prompt = *dparams.prompt;

-        if (sinfo.cache_size < prompt.size() + 1) {
-            llama_tokens tokens_new;
-            tokens_new.reserve(prompt.size() + 1 - sinfo.cache_size);
-            for (size_t j = sinfo.cache_size; j < prompt.size(); ++j) {
-                tokens_new.push_back(prompt[j]);
-            }
-            tokens_new.push_back(dparams.id_last); // add the last token
-
-            // Update context ngram cache with new dparams.prompt:
-            common_ngram_cache_update(
-                    sinfo.ngram_cache_context,
-                    LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX,
-                    tokens_new, tokens_new.size(), false);
-            sinfo.cache_size = prompt.size() + 1;
-        }
-
        llama_tokens inp;
        inp.reserve(prompt.size() + 1);
        for (size_t j = 0; j < prompt.size(); ++j) {
@@ -2249,6 +2306,16 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
        }
        inp.push_back(dparams.id_last);

+        if (sinfo.cache_size < inp.size()) {
+            // Update context ngram cache with the full history so n-grams
+            // ending at the new tokens have their preceding context
+            common_ngram_cache_update(
+                    sinfo.ngram_cache_context,
+                    LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX,
+                    inp, inp.size() - sinfo.cache_size, false);
+            sinfo.cache_size = inp.size();
+        }
+
        result.push_back(dparams.id_last);

        common_ngram_cache_draft(
@@ -2309,20 +2376,12 @@ static common_ngram_map get_common_ngram_map(
    return common_ngram_map(size_key, size_value, key_only, min_hits);
 }

-static common_speculative_impl_ngram_cache create_state_ngram_cache(
+static std::unique_ptr<common_speculative_impl_ngram_cache> create_state_ngram_cache(
        const common_speculative_config & config,
-        uint32_t n_seq,
-        const std::string & path_static,
-        const std::string & path_dynamic) {
+        uint32_t n_seq) {
    uint16_t n_draft = 8; // TODO get from config?

-    // TODO bool param in common/common.h to set save_static/save_dynamic?
-    bool save_static = false;
-    bool save_dynamic = false;
-
-    common_speculative_impl_ngram_cache state(config.params, n_seq, n_draft, path_static, path_dynamic, save_static, save_dynamic);
-
-    return state;
+    return std::make_unique<common_speculative_impl_ngram_cache>(config.params, n_seq, n_draft);
 }

 std::string common_speculative_type_name_str(const std::vector<common_speculative_type> & types) {
@@ -2822,11 +2881,7 @@ common_speculative * common_speculative_init(common_params_speculative & params,
                break;
            }
            case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: {
-                auto state = create_state_ngram_cache(
-                        config, n_seq,
-                        params.ngram_cache.lookup_cache_static,
-                        params.ngram_cache.lookup_cache_dynamic);
-                impls.push_back(std::make_unique<common_speculative_impl_ngram_cache>(state));
+                impls.push_back(create_state_ngram_cache(config, n_seq));
                break;
            }
            default:
@@ -73,7 +73,7 @@ struct common_speculative_draft_params {
    // the generated draft from the last _draft() call
    llama_tokens * result;

-    // candidate distribution per drafted token; set it to make draft-simple and draft-mtp sample
+    // candidate distribution per drafted token; set it to make draft-simple, draft-mtp and draft-dflash (DFlash2 only) sample
    std::vector<std::vector<llama_token_data>> * result_q = nullptr;

    // the target's temp and seed, read only when the drafter samples probabilistically
@@ -17,8 +17,10 @@ __all__ = [


 TEXT_MODEL_MAP: dict[str, str] = {
+    "NEOChatModel": "sensenova_u1",
    "AfmoeForCausalLM": "afmoe",
    "LagunaForCausalLM": "laguna",
+    "FlintForCausalLM": "flint",
    "ApertusForCausalLM": "llama",
    "ArceeForCausalLM": "llama",
    "ArcticForCausalLM": "arctic",
@@ -68,6 +70,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "LingDSparkModel": "qwen",
    "DeepseekV4ForCausalLM": "deepseek",
    "DeepseekV4DSparkModel": "deepseek",
+    "DeepseekV41ForCausalLM": "deepseek41",
    "DistilBertForMaskedLM": "bert",
    "DistilBertForSequenceClassification": "bert",
    "DistilBertModel": "bert",
@@ -194,13 +197,16 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "MiniCPMForCausalLM": "minicpm",
    "MiniCPMV4_6ForConditionalGeneration": "minicpm",
    "MiniCPMV4_7ForConditionalGeneration": "minicpm",
+    "MiniCPMOcrForConditionalGeneration": "minicpm",
    "MiniMaxText01ForCausalLM": "minimax",
    "MiniMaxM1ForCausalLM": "minimax",
    "MiniMaxM2ForCausalLM": "minimax",
    "MiniMaxM3SparseForCausalLM": "minimax",
    "MiniMaxM3SparseForConditionalGeneration": "minimax",
    "Ministral3ForCausalLM": "mistral3",
+    "Ministral3Model": "mistral3",
    "Mistral3ForConditionalGeneration": "mistral3",
+    "Mistral3Model": "mistral3",
    "VoxtralRealtimeForConditionalGeneration": "mistral3",
    "MistralForCausalLM": "llama",
    "MixtralForCausalLM": "llama",
@@ -226,6 +232,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "MuseSparkForConditionalGeneration": "muse_spark",
    "OpenELMForCausalLM": "openelm",
    "OpenJevModel": "qwen",
+    "H2OLightningModel": "h2o_lightning",
    "OrionForCausalLM": "orion",
    "PLMForCausalLM": "plm",
    "PLaMo2ForCausalLM": "plamo",
@@ -280,6 +287,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "SeedOssForCausalLM": "olmo",
    "SmallThinkerForCausalLM": "smallthinker",
    "SmolLM3ForCausalLM": "llama",
+    "SopranoTextModel": "soprano",
    "Spark2_5ForCausalLM": "spark2_5",
    "SolarOpenForCausalLM": "glm",
    "StableLMEpochForCausalLM": "stablelm",
@@ -319,6 +327,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "VllmD3Model": "vllm_d3",
    "DeepseekOCR2ForCausalLM": "deepseek",
    "DeepseekOCRForCausalLM": "deepseek",
+    "DeepseekV41ForCausalLM": "deepseek41",
    "DeepseekV4ForCausalLM": "deepseek",
    "Dots3NoteForCausalLM": "dots3",
    "Dots3NoteForConditionalGeneration": "dots3",
@@ -356,13 +365,16 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "MiniMaxM3SparseForConditionalGeneration": "minimax",
    "MiniCPMV4_6ForConditionalGeneration": "minicpm",
    "MiniCPMV4_7ForConditionalGeneration": "minicpm",
+    "MiniCPMOcrForConditionalGeneration": "minicpm",
    "Mistral3ForConditionalGeneration": "llava",
+    "Mistral3Model": "llava",
    "NemotronH_Omni_Reasoning_V3": "nemotron",
    "NemotronH_Nano_VL_V2": "nemotron",
    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
    "MuseSparkForConditionalGeneration": "muse_spark",
    "PaddleOCRVisionModel": "ernie",
    "Phi4ForCausalLMV": "phi",
+    "NEOVisionModel": "sensenova_u1",
    "Qwen2AudioForConditionalGeneration": "ultravox",
    "Qwen2VLForConditionalGeneration": "qwenvl",
    "Qwen2VLModel": "qwenvl",
@@ -371,10 +383,12 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "Qwen3ASRForConditionalGeneration": "qwen3vl",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
    "PocketTTSModel": "pockettts",
+    "SopranoModel": "soprano",
    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
    "OpenJevModel": "qwen3vl",
+    "H2OLightningModel": "h2o_lightning",
    "Qwen3_5ForConditionalGeneration": "qwen3vl",
    "Qwen3_5MoeForConditionalGeneration": "qwen3vl",
    "Qwen4ExpForConditionalGeneration": "qwen4exp",
@@ -1701,7 +1701,7 @@ class TextModel(ModelBase):
            self.gguf_writer.add_expert_group_used_count(n_group_used)
            logger.info(f"gguf: expert groups used count = {n_group_used}")

-        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func", "expert_selection_fn", "router_score_func"], optional=True)) is not None:
+        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func", "expert_selection_fn", "router_score_func", "router_score_function"], optional=True)) is not None:
            if score_func == "sigmoid":
                self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
            elif score_func == "softmax":
@@ -1837,6 +1837,9 @@ class TextModel(ModelBase):
        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
            res = "minerva-7b"
+        if chkhsh == "6e4d72c6e5260b90606e26a0931fa15e4b42ec6e97719d12919f2b1f93478262":
+            # ref: https://huggingface.co/domyn/Domyn-Small-v1.0
+            res = "domyn-small"
        if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664":
            # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct
            res = "hunyuan"
@@ -1939,6 +1942,9 @@ class TextModel(ModelBase):
        if chkhsh == "52df12b4c8d4176e7481aab4b6e8454d1fd0a210a04a574f6d4e067d10e23c3e":
            # ref: https://huggingface.co/CohereLabs/North-Mini-Code-1.0
            res = "cohere2moe"
+        if chkhsh == "d5420e601e979078c670bc503d9ac411c311fa76b76a76267c31e0d6ba816637":
+            # ref: https://huggingface.co/RecursiveSI/flint
+            res = "flint"
        if chkhsh == "e636dc30a262dcc0d8c323492e32ae2b70728f4df7dfe9737d9f920a282b8aea":
            # ref: https://huggingface.co/Qwen/Qwen1.5-7B
            res = "qwen2"
@@ -3143,7 +3149,7 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
    # Kimi-K3's text_config reports "KimiLinearForCausalLM", which is the older
    # Kimi-Linear-48B architecture and cannot load K3 (no attention residuals,
    # latent MoE, situ, ...). Route on the top-level architecture instead.
-    if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration"):
+    if model_type == ModelType.TEXT and arch in ("NEOChatModel", "StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration"):
        return arch

    # if "architectures" is found in the sub-config, use that instead
@@ -803,9 +803,6 @@ class DeepseekV4Model(TextModel):
        self._dsv4_mxfp4_generated = True
        return ()

-    def _format_dsv4_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None, suffix: str = ".weight") -> str:
-        return self.format_tensor_name(key, bid, suffix)
-
    def _map_dsv4_tensor_name(self, name: str, bid: int | None) -> tuple[gguf.MODEL_TENSOR, str]:
        root_map: dict[str, tuple[gguf.MODEL_TENSOR, str]] = {
            "embed.weight": (gguf.MODEL_TENSOR.TOKEN_EMBD, ".weight"),
@@ -879,17 +876,17 @@ class DeepseekV4Model(TextModel):

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if re.match(r"layers\.\d+\.ffn\.experts\.\d+\.w[123]\.(weight|scale)$", name):
-            return []
+            return

        # hash layers route text tokens via tid2eid and image tokens via bias_vl; gate.bias is unused
        if name.endswith(".ffn.gate.bias") and bid is not None and bid < self.hparams["num_hash_layers"]:
-            return []
+            return

        tensor_key, suffix = self._map_dsv4_tensor_name(name, bid)
        if tensor_key == gguf.MODEL_TENSOR.FFN_GATE_TID2EID:
-            return []
+            return

-        return [(self._format_dsv4_tensor_name(tensor_key, bid, suffix), data_torch)]
+        yield self.format_tensor_name(tensor_key, bid, suffix), data_torch

    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
        del bid  # unused
@@ -0,0 +1,464 @@
+from __future__ import annotations
+
+import json
+import math
+import re
+from pathlib import Path
+from typing import Any, Callable, Iterable, TYPE_CHECKING, cast
+
+import numpy as np
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import LazyTorchTensor, ModelBase, TextModel, gguf
+from .deepseek import DeepseekV4FlashVisionModel, DeepseekV4Model
+
+
+# Engram hash constants from inference/engram.py, stored in the GGUF: the token map needs the HF tokenizers
+# normalizers and the multipliers numpy's PCG64 stream, neither of which the runtime has.
+
+def _is_prime(n: int) -> bool:
+    if n < 2 or n % 2 == 0:
+        return n == 2
+    return all(n % d for d in range(3, math.isqrt(n) + 1, 2))
+
+
+def engram_primes(n_layers: int, max_ngram_size: int, n_heads: int, vocab_size: int) -> list[list[list[int]]]:
+    """EngramLayout.from_args: [layer][ngram order 2..max][head] bucket moduli, never reused across layers."""
+    seen: set[int] = set()
+    out: list[list[list[int]]] = []
+    for _ in range(n_layers):
+        per_ngram = []
+        for _ in range(max_ngram_size - 1):
+            sizes, cur = [], vocab_size - 1
+            for _ in range(n_heads):
+                cand = cur + 1
+                while not _is_prime(cand) or cand in seen:
+                    cand += 1
+                cur = cand
+                seen.add(cur)
+                sizes.append(cur)
+            per_ngram.append(sizes)
+        out.append(per_ngram)
+    return out
+
+
+def engram_multipliers(layer_ids: list[int], max_ngram_size: int, compressed_vocab_size: int) -> list[list[int]]:
+    """compute_hash_multipliers: odd, bounded so that token_id * multiplier cannot overflow int64."""
+    bound = max(1, (np.iinfo(np.int64).max // compressed_vocab_size) // 2)
+    rows = []
+    for layer_id in layer_ids:
+        rng = np.random.default_rng(10007 * layer_id)
+        values = rng.integers(low=0, high=bound, size=(max_ngram_size,), dtype=np.int64)
+        rows.append([int(v) * 2 + 1 for v in values])
+    return rows
+
+
+def engram_compressed_token_map(tokenizer_dir: Path) -> tuple[list[int], int]:
+    """build_compressed_token_map: token id -> compressed id (tokens that normalize alike collapse)."""
+    from tokenizers import Regex, normalizers
+    from transformers import AutoTokenizer
+
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir)
+    sentinel = "\ue000"
+    normalizer = normalizers.Sequence([
+        normalizers.NFKC(),
+        normalizers.NFD(),
+        normalizers.StripAccents(),
+        normalizers.Lowercase(),
+        normalizers.Replace(Regex(r"[ \t\r\n]+"), " "),
+        normalizers.Replace(Regex(r"^ $"), sentinel),
+        normalizers.Strip(),
+        normalizers.Replace(sentinel, " "),
+    ])
+    backend = tokenizer.backend_tokenizer
+    key_to_new: dict[str, int] = {}
+    lookup = [0] * len(tokenizer)
+    for token_id in range(len(tokenizer)):
+        text = backend.decode([token_id], skip_special_tokens=False)
+        if "\ufffd" in text:
+            key = backend.id_to_token(token_id)
+        else:
+            normalized = normalizer.normalize_str(text)
+            key = normalized if normalized else text
+        lookup[token_id] = key_to_new.setdefault(key, len(key_to_new))
+    return lookup, len(key_to_new)
+
+
+class _StreamedBytes:
+    """Raw tensor bytes produced chunk by chunk at write time (an Engram table is ~200 GB in BF16)."""
+    def __init__(self, shape: tuple[int, ...], chunks: Callable[[], Iterable[np.ndarray]]):
+        self.shape = shape
+        self.dtype = np.dtype(np.uint8)
+        self.nbytes = int(np.prod(shape))
+        self._chunks = chunks
+
+    def tofile(self, fout) -> None:
+        written = 0
+        for chunk in self._chunks():
+            chunk = np.ascontiguousarray(chunk, dtype=np.uint8)
+            chunk.tofile(fout)
+            written += chunk.nbytes
+        if written != self.nbytes:
+            raise ValueError(f"streamed {written} bytes, expected {self.nbytes}")
+
+    def byteswap(self, inplace: bool = False):
+        raise NotImplementedError("streamed tensors are written little-endian only")
+
+
+def _e8m0_check(scale_u8: np.ndarray, name: str) -> None:
+    # e4m3 (|x| in [2^-9, 448]) * 2^(e - 127) is exact in BF16 for e in [3, 246]
+    if scale_u8.size and (int(scale_u8.min()) < 3 or int(scale_u8.max()) > 246):
+        raise ValueError(f"{name}: E8M0 scale exponents outside [3, 246], the dequantized values would not be exact")
+
+
+@ModelBase.register("DeepseekV41ForCausalLM")
+class DeepseekV41Model(DeepseekV4Model):
+    """DeepSeek-V4.1 text model, lossless: FP8 (e4m3, 32x32 ue8m0 blocks) linears dequantized to BF16 (F32 with
+    --outtype f32), routed experts copied bit-exact as MXFP4, Engram tables (e4m3 rows, one e8m0 scale per 32) as the
+    exact BF16 lookup (F32 with --outtype f32), router and Engram gate F32. DSpark (mtp.*) is not converted here
+    (--dspark, DeepseekV41DSparkModel); the vision tower goes to an mmproj (--mmproj, DeepseekV41VisionModel).
+    The Engram tables of all layers are written as one per_layer_token_embd (rows concatenated in layer order) with
+    the PLE keys of qwen4exp: absolute head offsets and bucket counts, so the runtime can read it lazily."""
+    model_arch = gguf.MODEL_ARCH.DEEPSEEK41
+    supports_mtp_export = False
+
+    _ENGRAM_EMBED_RE = re.compile(r"layers\.(\d+)\.engram\.embed\.(weight|scale)$")
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams.setdefault("num_hash_layers", 0)  # no hash-routed layers in V4.1
+        self._f32 = self.ftype == gguf.LlamaFileType.ALL_F32
+        qc = self.hparams.get("quantization_config") or {}
+        self._fp8_block = tuple(qc.get("weight_block_size") or (32, 32))
+        if qc.get("scale_fmt", "ue8m0") != "ue8m0":
+            raise ValueError(f"unsupported FP8 scale format {qc.get('scale_fmt')!r}")
+        if self.hparams.get("gate_temp", 1.0) != 1.0:
+            raise ValueError("gate_temp != 1 is not supported (the router logits would need dividing by it)")
+
+        # V4.1 does not use the DeepSeek-V4 template; a template shipped with the model is still added by set_vocab
+        self.gguf_writer.remove_key(gguf.Keys.Tokenizer.CHAT_TEMPLATE)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        # num_hidden_layers lives in text_config, which is merged only after indexing; MTP stages are skipped anyway
+        return TextModel.index_tensors(self, remote_hf_model_id=remote_hf_model_id)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hp = self.hparams
+        w = self.gguf_writer
+        w.add_attention_kv_source_layers(hp["kv_source_layer_ids"])
+        w.add_indexer_source_layers(hp["index_source_layer_ids"])
+        if hp.get("candidate_source_layer_id", -1) >= 0:
+            w.add_indexer_candidate_source_layer(hp["candidate_source_layer_id"])
+            w.add_indexer_candidate_top_k_blocks(hp["candidate_topk_blocks"])
+            w.add_indexer_candidate_block_size(hp["candidate_block_size"])
+
+        # GGUF cannot hold empty arrays: the Engram keys are omitted when the model has no Engram layers
+        layer_ids = list(hp.get("engram_layer_ids") or [])
+        if layer_ids:
+            max_ngram, n_heads = hp["engram_max_ngram_size"], hp["engram_n_heads"]
+            primes = engram_primes(len(layer_ids), max_ngram, n_heads, hp["engram_vocab_size"])
+            for li, layer_primes in enumerate(primes):
+                if sum(p for per in layer_primes for p in per) != hp["engram_num_embeddings"][li]:
+                    raise ValueError(f"engram layer {layer_ids[li]}: prime sum != num_embeddings")
+            token_map, cvocab = engram_compressed_token_map(self.dir_model)
+            if cvocab != hp["engram_compressed_vocab_size"] or len(token_map) != hp["vocab_size"]:
+                raise ValueError(f"compressed token map ({len(token_map)} -> {cvocab}) does not match the config")
+            # absolute row offsets into the concatenated table
+            sizes = [p for layer in primes for per in layer for p in per]
+            offsets = [sum(sizes[:i]) for i in range(len(sizes))]
+            w.add_key_value(gguf.Keys.PerLayerEmbedding.LAYERS.format(arch=w.arch), layer_ids,
+                            gguf.GGUFValueType.ARRAY, gguf.GGUFValueType.UINT32)
+            w.add_ple_ngram_size(max_ngram)
+            w.add_ple_heads_per_ngram(n_heads)
+            w.add_embedding_length_per_layer_input(hp["engram_head_dim"])
+            w.add_ple_layer_multipliers([m for row in engram_multipliers(layer_ids, max_ngram, cvocab) for m in row])
+            w.add_ple_head_offsets(offsets)
+            w.add_ple_head_vocab_sizes(sizes)
+            w.add_engram_pad_token_id(hp["engram_pad_token_id"])
+            w.add_engram_token_map(token_map)
+
+    def dequant_model(self):
+        bo, bi = self._fp8_block
+
+        def dequant(weight: Tensor, scale: Tensor) -> Tensor:
+            out_f, in_f = weight.shape
+            s = self._e8m0_to_float(scale).repeat_interleave(bo, 0)[:out_f].repeat_interleave(bi, 1)[:, :in_f]
+            return weight.float() * s
+
+        for name in [n for n in self.model_tensors if n.endswith(".scale") and not self._ENGRAM_EMBED_RE.match(n)]:
+            weight_name = name.removesuffix(".scale") + ".weight"
+            if weight_name not in self.model_tensors or self.model_tensors[weight_name]().dtype not in self._float8_dtypes():
+                continue
+            w, s = self.model_tensors[weight_name], self.model_tensors.pop(name)
+            out_f, in_f = w().shape
+            if tuple(s().shape) != (-(-out_f // bo), -(-in_f // bi)):
+                raise ValueError(f"{weight_name}: scale shape {tuple(s().shape)} does not match {(bo, bi)} blocks")
+            _e8m0_check(LazyTorchTensor.to_eager(s()).view(torch.uint8).numpy(), name)
+            self.model_tensors[weight_name] = lambda w=w, s=s: dequant(w(), s())
+            self._dsv4_fp8_dequantized.add(weight_name)
+
+    def _tensor_part(self, name: str) -> Path:
+        index = self.dir_model / "model.safetensors.index.json"
+        if not index.is_file():
+            return self.dir_model / "model.safetensors"
+        with open(index, "r", encoding="utf-8") as f:
+            return self.dir_model / json.load(f)["weight_map"][name]
+
+    def _write_mxfp4_experts(self, bid: int, proj: str, key: gguf.MODEL_TENSOR, n_expert: int | None = None) -> None:
+        n_expert = self.hparams["n_routed_experts"] if n_expert is None else n_expert
+        names = [f"layers.{bid}.ffn.experts.{e}.{proj}" for e in range(n_expert)]
+        pairs = [(self.model_tensors.pop(n + ".weight"), self.model_tensors.pop(n + ".scale")) for n in names]
+        rows, packed_cols = pairs[0][0]().shape
+
+        def chunks() -> Iterable[np.ndarray]:
+            for w, s in pairs:
+                yield self.repack_mxfp4_blocks(LazyTorchTensor.to_eager(w()), LazyTorchTensor.to_eager(s()))
+        data = cast(np.ndarray, _StreamedBytes((len(pairs), rows, packed_cols * 2 // 32 * 17), chunks))
+        self.gguf_writer.add_tensor(self.format_tensor_name(key, bid), data, raw_dtype=gguf.GGMLQuantizationType.MXFP4)
+
+    def _write_engram_tables(self, layer_ids: list[int]) -> None:
+        # official lookup e4m3 * 2^(e8m0 - 127), exact in BF16
+        parts, dim = [], None
+        for bid in layer_ids:
+            wn, sn = f"layers.{bid}.engram.embed.weight", f"layers.{bid}.engram.embed.scale"
+            rows, d = self.model_tensors.pop(wn)().shape
+            if d % 32 != 0 or tuple(self.model_tensors.pop(sn)().shape) != (rows, d // 32) or d != (dim or d):
+                raise ValueError(f"{wn}: expected one e8m0 scale per 32 elements and the same width in every layer")
+            if rows != self.hparams["engram_num_embeddings"][len(parts)]:
+                raise ValueError(f"{wn}: {rows} rows, config says {self.hparams['engram_num_embeddings'][len(parts)]}")
+            dim = d
+            parts.append((wn, sn, rows, self._tensor_part(wn), self._tensor_part(sn)))
+        assert dim is not None
+        name = self.format_tensor_name(gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD)
+        n_rows = sum(p[2] for p in parts)
+
+        def chunks() -> Iterable[np.ndarray]:
+            from safetensors import safe_open
+            for wn, sn, rows, wpart, spart in parts:
+                with safe_open(wpart, framework="pt") as fw, safe_open(spart, framework="pt") as fs:
+                    ws, ss = fw.get_slice(wn), fs.get_slice(sn)
+                    for r0 in range(0, rows, 1 << 18):
+                        q = ws[r0:r0 + (1 << 18)].view(torch.uint8).reshape(-1, 32)
+                        e = ss[r0:r0 + (1 << 18)].view(torch.uint8).reshape(-1, 1)
+                        _e8m0_check(e.numpy(), sn)
+                        v = (q.view(torch.float8_e4m3fn).float() * torch.exp2(e.float() - 127)).reshape(-1, dim)
+                        yield (v if self._f32 else v.bfloat16()).view(torch.uint8).numpy()
+        qt = gguf.GGMLQuantizationType.F32 if self._f32 else gguf.GGMLQuantizationType.BF16
+        data = cast(np.ndarray, _StreamedBytes((n_rows, dim * (4 if self._f32 else 2)), chunks))
+        self.gguf_writer.add_tensor(name, data, raw_dtype=qt)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        if not self._dsv4_mxfp4_generated:
+            for bid in range(self.block_count):
+                self._write_mxfp4_experts(bid, "w1", gguf.MODEL_TENSOR.FFN_GATE_EXP)
+                self._write_mxfp4_experts(bid, "w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP)
+                self._write_mxfp4_experts(bid, "w3", gguf.MODEL_TENSOR.FFN_UP_EXP)
+            if self.hparams.get("engram_layer_ids"):
+                self._write_engram_tables(list(self.hparams["engram_layer_ids"]))
+            self._dsv4_mxfp4_generated = True
+        return ()
+
+    _LAYER_MAP: dict[str, tuple[gguf.MODEL_TENSOR, str]] = {
+        "attn.indexer.wk.weight": (gguf.MODEL_TENSOR.INDEXER_ATTN_K, ".weight"),
+        "attn.indexer.k_norm.weight": (gguf.MODEL_TENSOR.INDEXER_K_NORM, ".weight"),
+        "engram.wkv.weight": (gguf.MODEL_TENSOR.ENGRAM_KV, ".weight"),
+    }
+
+    def _map_dsv4_tensor_name(self, name: str, bid: int | None) -> tuple[gguf.MODEL_TENSOR, str]:
+        m = re.match(r"layers\.\d+\.(.+)$", name)
+        if m is not None and m.group(1) in self._LAYER_MAP:
+            return self._LAYER_MAP[m.group(1)]
+        return super()._map_dsv4_tensor_name(name, bid)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith(".engram.k_weight"):
+            return  # folded into engram_gate
+        if name.endswith(".engram.q_weight"):
+            # Engram gate weight = q_weight * k_weight (bf16 * bf16 is exact in f32)
+            k = self.model_tensors[name.removesuffix("q_weight") + "k_weight"]
+            yield self.format_tensor_name(gguf.MODEL_TENSOR.ENGRAM_GATE, bid), data_torch.float() * k().float()
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
+        if new_name.endswith((".engram_gate.weight", ".ffn_gate_inp.weight")) or name in self._dsv4_f32_tensors:
+            return gguf.GGMLQuantizationType.F32
+        if n_dims >= 2 and (name in self._dsv4_fp8_dequantized or name in self._dsv4_bf16_tensors):
+            return gguf.GGMLQuantizationType.F32 if self._f32 else gguf.GGMLQuantizationType.BF16
+        return False
+
+
+class DeepseekV41DSparkModel(DeepseekV41Model):
+    """DeepSeek-V4.1 DSpark speculative draft (mtp.0..2 of the checkpoint) as a DFLASH GGUF (`--dspark`).
+
+    Layout = the V4 DSpark draft (conversion/deepseek.py DeepseekV4DSparkModel) with V4.1's differences flagged by
+    `dflash.hyper_connection.single_pass`: stages blk.0..2 (MXFP4 experts bit-exact, dense BF16), mtp.0.main_proj /
+    main_norm -> fc / enc.output_norm, mtp.2.norm -> output_norm, markov_head.embed / head -> markov_w1 / markov_w2,
+    confidence_head.proj -> conf_proj (F32), no hc_head. Token embeddings and the lm head are the target's (shared at
+    runtime through ctx_other). target_layers are the V4.1 ids as-is: the draft reads the attention input of those
+    layers, which is what llama_set_embeddings_layer_inp extracts.
+    """
+    model_arch = gguf.MODEL_ARCH.DFLASH
+    mtp_only = True
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hp = self.hparams
+        self._n_main = hp["num_hidden_layers"]
+        self._n_stage = hp["num_nextn_predict_layers"]
+        if self._n_stage <= 0 or "dspark_block_size" not in hp:
+            raise ValueError("DeepSeek-V4.1 DSpark export needs num_nextn_predict_layers > 0 and the dspark_* config")
+        # stages keep main-model block ids (mtp.M -> layers.n_main+M); format_tensor_name() maps them back
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self._n_stage)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        hp = {**self.hparams, **self.hparams.get("text_config", {})}  # text_config is merged only after indexing
+        type(self)._dsv4_main_layers = hp["num_hidden_layers"]
+        type(self)._dsv4_nextn_layers = hp.get("num_nextn_predict_layers", 0)
+        return TextModel.index_tensors(self, remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        m = re.match(r"mtp\.(\d+)\.(.+)$", name)
+        if m is None:
+            return None  # shared with the target at runtime
+        stage, rest = int(m.group(1)), m.group(2)
+        assert cls._dsv4_main_layers is not None
+        if stage >= cls._dsv4_nextn_layers:
+            raise ValueError(f"Unexpected DSpark stage {stage}")
+        if rest.startswith(("markov_head.", "confidence_head.")):
+            return f"dspark.{rest}", gen
+        return f"layers.{cls._dsv4_main_layers + stage}.{rest}", gen
+
+    def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
+        if bid is not None and bid >= self._n_main:
+            bid -= self._n_main
+        return super().format_tensor_name(key, bid, suffix)
+
+    def set_vocab(self):
+        super().set_vocab()
+        self.gguf_writer.add_mask_token_id(self.hparams["dspark_noise_token_id"])
+
+    def set_gguf_parameters(self):
+        hp = self.hparams
+        n = self._n_stage
+        w = self.gguf_writer
+        w.add_block_count(n)
+        w.add_context_length(hp["max_position_embeddings"])
+        w.add_embedding_length(hp["hidden_size"])
+        w.add_feed_forward_length(hp["moe_intermediate_size"])
+        w.add_head_count(hp["num_attention_heads"])
+        w.add_head_count_kv(1)
+        w.add_key_length(hp["head_dim"])
+        w.add_value_length(hp["head_dim"])
+        w.add_rope_dimension_count(hp["qk_rope_head_dim"])
+        # compress-ratio-0 layers: plain rope at rope_theta, no YaRN
+        w.add_rope_freq_base(float(hp["rope_theta"]))
+        w.add_layer_norm_rms_eps(hp["rms_norm_eps"])
+        w.add_q_lora_rank(hp["q_lora_rank"])
+        w.add_sliding_window(hp["sliding_window"])
+        w.add_expert_count(hp["dspark_n_routed_experts"])
+        w.add_expert_used_count(hp["dspark_num_experts_per_tok"])
+        w.add_expert_feed_forward_length(hp["moe_intermediate_size"])
+        w.add_expert_shared_count(hp["n_shared_experts"])
+        w.add_expert_weights_scale(hp["routed_scaling_factor"])
+        w.add_expert_weights_norm(hp["norm_topk_prob"])
+        w.add_expert_gating_func(gguf.ExpertGatingFuncType.SQRTSOFTPLUS)
+        w.add_swiglu_clamp_exp([hp["swiglu_limit"]] * n)
+        w.add_swiglu_clamp_shexp([hp["swiglu_limit"]] * n)
+        w.add_attention_output_group_count(hp["o_groups"])
+        w.add_attention_output_lora_rank(hp["o_lora_rank"])
+        w.add_attention_compress_ratios([0] * n)
+        w.add_hyper_connection_count(hp["hc_mult"])
+        w.add_hyper_connection_sinkhorn_iterations(hp["hc_sinkhorn_iters"])
+        w.add_hyper_connection_epsilon(hp["hc_eps"])
+        w.add_hyper_connection_single_pass(True)
+        w.add_block_size(hp["dspark_block_size"])
+        w.add_target_layers(hp["dspark_target_layer_ids"])
+        w.add_sample_from_anchor(True)
+        w.add_has_confidence_head(True)
+        w.add_file_type(self.ftype)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # no Engram tables in the draft
+        if not self._dsv4_mxfp4_generated:
+            n_expert = self.hparams["dspark_n_routed_experts"]
+            for bid in range(self._n_main, self._n_main + self._n_stage):
+                self._write_mxfp4_experts(bid, "w1", gguf.MODEL_TENSOR.FFN_GATE_EXP, n_expert)
+                self._write_mxfp4_experts(bid, "w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP, n_expert)
+                self._write_mxfp4_experts(bid, "w3", gguf.MODEL_TENSOR.FFN_UP_EXP, n_expert)
+            self._dsv4_mxfp4_generated = True
+        return ()
+
+    _DSPARK_ROOT: dict[str, gguf.MODEL_TENSOR] = {
+        "main_proj.weight": gguf.MODEL_TENSOR.FC,
+        "main_norm.weight": gguf.MODEL_TENSOR.ENC_OUTPUT_NORM,
+        "norm.weight": gguf.MODEL_TENSOR.OUTPUT_NORM,
+        "dspark.markov_head.embed.weight": gguf.MODEL_TENSOR.DSPARK_MARKOV_W1,
+        "dspark.markov_head.head.weight": gguf.MODEL_TENSOR.DSPARK_MARKOV_W2,
+        "dspark.confidence_head.proj.weight": gguf.MODEL_TENSOR.DSPARK_CONF_PROJ,
+    }
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith(".ffn.gate.bias_vl"):
+            return
+        if name in self._DSPARK_ROOT:
+            yield self.format_tensor_name(self._DSPARK_ROOT[name], None), data_torch
+            return
+        m = re.match(r"layers\.\d+\.(main_proj\.weight|main_norm\.weight|norm\.weight)$", name)
+        if m is not None:
+            yield self.format_tensor_name(self._DSPARK_ROOT[m.group(1)], None), data_torch
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
+        if new_name == "conf_proj.weight":
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+
+@ModelBase.register("DeepseekV41ForCausalLM")
+@ModelBase.example("deepseek-ai/DeepSeek-V4.1-Flash-1017")
+class DeepseekV41VisionModel(DeepseekV4FlashVisionModel):
+    """DeepSeek-V4.1 vision tower (projector deepseek41v): the V4-Flash-Vision ViT + 3x3 aligner, with the V4.1 token
+    block ([START] ([IMAGE]*w [NEWLINE])*h [END], no pads) and resize plan. ref: inference/vision.py,
+    inference/image_processor.py of deepseek-ai/DeepSeek-V4.1-Flash-1017. The checkpoint config is HF-style
+    (vision_config); image_start / image_end / image_newline are written like V4's sentinels, there is no image_pad."""
+
+    def get_vision_config(self) -> dict[str, Any] | None:
+        vc = self.global_config.get("vision_config")
+        if not vc or vc.get("num_hidden_layers", 0) == 0:
+            raise ValueError("DeepseekV41VisionModel requires vision_config.num_hidden_layers > 0")
+        return {
+            "num_hidden_layers":   vc["num_hidden_layers"],
+            "hidden_size":         vc["hidden_size"],
+            "num_attention_heads": vc["num_attention_heads"],
+            "intermediate_size":   vc["intermediate_size"],
+            "patch_size":          vc["patch_size"],
+            # dynamic resolution; only used for compat / warmup
+            "image_size":          vc["patch_size"] * vc["downsample_ratio"] * 16,
+            "rope_theta":          vc.get("rope_theta", 10000.0),
+            "downsample_ratio":    vc["downsample_ratio"],
+            "min_pixels":          vc["min_pixels"],
+            "max_image_tokens":    vc["max_image_tokens"],
+            "max_wh_ratio":        vc.get("max_wh_ratio"),
+        }
+
+    def set_gguf_parameters(self):
+        assert self.hparams_vision is not None
+        super(DeepseekV4FlashVisionModel, self).set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DEEPSEEK41V)
+        # pytorch default eps (vision.py), not the LLM rms_norm_eps (1e-20)
+        self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
+        self.gguf_writer.add_vision_use_silu(True)
+        self.gguf_writer.add_vision_projector_scale_factor(self.hparams_vision["downsample_ratio"])
+        self.gguf_writer.add_vision_min_pixels(self.hparams_vision["min_pixels"])
+        # hardcoded in clip.cpp (PROJECTOR_TYPE_DEEPSEEK41V); add GGUF keys if a later model differs
+        assert self.hparams_vision["max_image_tokens"] == 1024, self.hparams_vision["max_image_tokens"]
+        assert self.hparams_vision["max_wh_ratio"] is None, self.hparams_vision["max_wh_ratio"]
+        assert self.hparams_vision["rope_theta"] == 10000, self.hparams_vision["rope_theta"]
@@ -0,0 +1,38 @@
+from __future__ import annotations
+
+from .base import ModelBase, gguf
+from .qwen import Qwen2MoeModel
+
+
+@ModelBase.register("FlintForCausalLM")
+@ModelBase.example("RecursiveSI/flint")
+class FlintModel(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.FLINT
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        dense = self.hparams["mlp_only_layers"]
+        if dense != list(range(len(dense))):
+            raise ValueError("Flint dense layers must precede the MoE layers")
+        self.gguf_writer.add_leading_dense_block_count(len(dense))
+        self.gguf_writer.add_expert_shared_count(
+            self.hparams["shared_expert_intermediate_size"] // self.hparams["moe_intermediate_size"]
+        )
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+        self.gguf_writer.add_expert_weights_scale(1.0)
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern([
+            kind == "sliding_attention" for kind in self.hparams["layer_types"]
+        ])
+        self.gguf_writer.add_rope_pattern([
+            i not in self.hparams["no_rope_layer_indices"] for i in range(self.block_count)
+        ])
+
+    def modify_tensors(self, data_torch, name, bid):
+        if name.endswith(".softmax_offset"):
+            name += ".weight"
+        if name.endswith(".gate.expert_bias"):
+            if not self.hparams["router_expert_bias"]:
+                return
+            name += ".bias"
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -999,6 +999,8 @@ class Gemma4VisionAudioModel(MmprojModel):
        assert self.hparams_vision is not None
        self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.GEMMA4V)
        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
+        assert self.hparams_vision.get("hidden_activation", "gelu_pytorch_tanh") == "gelu_pytorch_tanh"
+        self.gguf_writer.add_vision_use_gelu(True)

        # audio params
        if self.has_audio_encoder:
@@ -0,0 +1,96 @@
+from __future__ import annotations
+
+import json
+
+from pathlib import Path
+from typing import Any
+
+from .base import ModelBase, gguf, logger
+from .qwen import Qwen3_5TextModel
+from .qwen3vl import Qwen3VLVisionModel
+
+
+def _is_h2o_lightning_checkpoint(dir_model: Path) -> bool:
+    return (dir_model / "h2o_lightning_shim.py").is_file() and (dir_model / "serve_config.json").is_file()
+
+
+@ModelBase.register_hparams_loader(_is_h2o_lightning_checkpoint)
+def _load_h2o_lightning_hparams(dir_model: Path) -> dict[str, Any]:
+    logger.info("gguf: detected H2O-Lightning checkpoint")
+    hparams = ModelBase.load_hparams(dir_model, False, guess=False)
+    hparams["architectures"] = ["H2OLightningModel"]
+    return hparams
+
+
+@ModelBase.register("H2OLightningModel")
+@ModelBase.example("h2oai/h2o-lightning-4b")
+class H2OLightningModel(Qwen3_5TextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN35
+
+    # prompt and calibration follow serve_config.json and h2o_lightning_shim.py of the model repo
+    # the state budget is hard-coded in server-decision.cpp
+    _MAX_STATE_TOKENS = 32000
+    _DEFAULT_INSTRUCTIONS = "Answer the question below."
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        with open(self.dir_model / "serve_config.json", "r", encoding="utf-8") as f:
+            self._serve_config = json.load(f)
+        prompt = self._serve_config["prompt"]
+        expected = {
+            "state_compact":    (prompt.get("state_compact", True), True),
+            "max_state_tokens": (prompt.get("max_state_tokens"), self._MAX_STATE_TOKENS),
+            "layout":           (prompt.get("layout", "plain"), "plain"),
+            "chat_template":    (prompt.get("chat_template", True), True),
+            "enable_thinking":  (prompt.get("enable_thinking", False), False),
+        }
+        for name, (got, want) in expected.items():
+            if got != want:
+                raise ValueError(f"serve_config.json: {name} is {got!r}, the h2o-lightning decision type expects {want!r}")
+
+    def get_vocab_base_pre(self, tokenizer) -> str:
+        # transformers loads this checkpoint as Qwen2Tokenizer: the model was trained and is served with the Qwen2 split
+        pre = super().get_vocab_base_pre(tokenizer)
+        if pre != "qwen35":
+            raise ValueError(f"expected the Qwen3.5 tokenizer, got pre-tokenizer {pre!r}")
+        return "qwen2"
+
+    def set_vocab(self):
+        super().set_vocab()
+        self.gguf_writer.add_chat_template([{"name": "systemone", "template": self._systemone_template()}])
+
+    def _systemone_template(self) -> str:
+        # the server gives the template texts rendered as the shim renders them, it only lays them out
+        prompt = self._serve_config["prompt"]
+        nl = "{{ '\\n' }}"
+        description = (
+            "{% if type == 'noul' %}"
+            "{% if o.description %}{{ o.description }}{% elif o.key == 'true' %}the statement holds{% else %}it does not{% endif %}"
+            "{% elif o.description is none %}{{ o.key }}{% else %}{{ o.description }}{% endif %}"
+        )
+        user = (
+            "{% set user %}record: {{ state }}" + nl
+            + "question: {{ instructions if instructions else '" + self._DEFAULT_INSTRUCTIONS + "' }}" + nl
+            + "options:{% for o in options %}" + nl + "{{ o.label | trim }}) {{ o.key }}: " + description + "{% endfor %}{% endset %}"
+        )
+        return (
+            "<|im_start|>system" + nl + prompt["system"].strip() + "<|im_end|>" + nl
+            + "<|im_start|>user" + nl + "{% for image in images %}{{ image }}{% endfor %}" + user + "{{ user | trim }}<|im_end|>" + nl
+            + "<|im_start|>assistant" + nl + "<think>{{ '\\n\\n' }}</think>{{ '\\n\\n' }}" + prompt["prefill"]
+        )
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_decision_type(gguf.DecisionType.H2O_LIGHTNING)
+        # the answer follows "Answer:", so the token of a label has a leading space
+        cfg = self._serve_config
+        self.gguf_writer.add_decision_labels([" " + label for label in cfg["labels"][:cfg.get("max_options", len(cfg["labels"]))]])
+        for name in ("choice", "noul", "score"):
+            self.gguf_writer.add_decision_temperature(name, self._serve_config["temperature"][name])
+        # a request with images has this temperature for all question types
+        self.gguf_writer.add_decision_temperature("image", self._serve_config["image"]["temperature"])
+
+
+@ModelBase.register("H2OLightningModel")
+class H2OLightningVisionModel(Qwen3VLVisionModel):
+    pass
@@ -374,6 +374,16 @@ class LFM2VLModel(MmprojModel):
            if resize_algo is None:
                raise ValueError(f"unsupported resample: {resample}")
            self.gguf_writer.add_vision_image_resize_algo(resize_algo)
+        # tiling params from processor_config.json (do_image_splitting=False -> single tile, i.e. no tiling)
+        preproc = self.preprocessor_config
+        if preproc.get("do_image_splitting", True):
+            self.gguf_writer.add_vision_preproc_min_tiles(preproc.get("min_tiles", 2))
+            self.gguf_writer.add_vision_preproc_max_tiles(preproc.get("max_tiles", 10))
+        else:
+            self.gguf_writer.add_vision_preproc_min_tiles(1)
+            self.gguf_writer.add_vision_preproc_max_tiles(1)
+        self.gguf_writer.add_vision_preproc_tile_size(preproc.get("tile_size", 512))
+        self.gguf_writer.add_vision_preproc_max_pixels_tolerance(preproc.get("max_pixels_tolerance", 2.0))

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
@@ -2,7 +2,7 @@ from __future__ import annotations

 import json

-from typing import Iterable, TYPE_CHECKING
+from typing import Callable, Iterable, TYPE_CHECKING

 if TYPE_CHECKING:
    from torch import Tensor
@@ -128,3 +128,24 @@ class LlavaVisionModel(MmprojModel):
            yield from super().modify_tensors(img_break_embd, name, bid)

        return # skip other tensors
+
+
+@ModelBase.register("Mistral3Model")
+@ModelBase.example("nvidia/Nemotron-3.5-Embed-1B-BF16")
+class Mistral3EmbeddingVisionModel(LlavaVisionModel):
+    def __init__(self, dir_model, *args, **kwargs):
+        hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        text_config = hparams.get("text_config", hparams)
+        if text_config.get("is_causal") is not False:
+            raise NotImplementedError("Expected a Nemotron Embed model for Mistral3Model, but is_causal is not False")
+        super().__init__(dir_model, *args, **kwargs)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if item[0].endswith("embed_tokens.weight"):
+            return ModelBase.filter_tensors(item)
+        return super().filter_tensors(item)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vision_preproc_image_size(self.preprocessor_config["size"]["longest_edge"])
@@ -272,3 +272,18 @@ class MiniCPMV4_7VisionModel(MiniCPMV4_6VisionModel):
            yield self.format_tensor_name(gguf.MODEL_TENSOR.V_TOK_EMBD_SEP, suffix=""), data_torch[ids]
            return
        yield from super().modify_tensors(data_torch, name, bid)
+
+
+# MiniCPM-OCR 1.0: the 4.6 stack with model_type minicpm_ocr.
+
+@ModelBase.register("MiniCPMOcrForConditionalGeneration")
+@ModelBase.example("openbmb/MiniCPM-OCR-1.0")
+class MiniCPMOCR1_0TextModel(MiniCPMV4_6TextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN35
+    no_mtp = True
+
+
+@ModelBase.register("MiniCPMOcrForConditionalGeneration")
+@ModelBase.example("openbmb/MiniCPM-OCR-1.0")
+class MiniCPMOCR1_0VisionModel(MiniCPMV4_6VisionModel):
+    projector_type = gguf.VisionProjectorType.MINICPM_OCR_1_0
@@ -5,7 +5,7 @@ from typing import TYPE_CHECKING
 if TYPE_CHECKING:
    from torch import Tensor

-from .base import ModelBase, TextModel, gguf
+from .base import ModelBase, TextModel, get_model_architecture, gguf

 from .deepseek import DeepseekV2Model
 from .llama import LlamaModel
@@ -13,9 +13,11 @@ from .llama import LlamaModel

@ModelBase.register(
    "Mistral3ForConditionalGeneration",
+    "Mistral3Model",
    "Ministral3ForCausalLM",
+    "Ministral3Model",
 )
-@ModelBase.example("mistralai/Mistral-Small-3.1-24B-Instruct-2503", "hf-tiny-v2/tiny-random-Ministral3ForCausalLM")
+@ModelBase.example("mistralai/Mistral-Small-3.1-24B-Instruct-2503", "hf-tiny-v2/tiny-random-Ministral3ForCausalLM", "nvidia/Nemotron-3-Embed-1B-BF16", "nvidia/Nemotron-3.5-Embed-1B-BF16")
 class Mistral3Model(TextModel):
    class Ministral3Model(LlamaModel):
        model_arch = gguf.MODEL_ARCH.MISTRAL3
@@ -29,6 +31,23 @@ class Mistral3Model(TextModel):
                self.gguf_writer.add_rope_scaling_yarn_log_mul(rope_params["mscale_all_dim"])
                self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])

+    class Mistral3EmbedNemotronModel(Ministral3Model):
+        model_arch = gguf.MODEL_ARCH.MISTRAL3
+
+        def set_vocab(self):
+            super().set_vocab()
+            self.gguf_writer.add_add_bos_token(False)
+            self.gguf_writer.add_add_eos_token(False)
+
+        def get_vocab_base_pre(self, tokenizer) -> str:
+            if self.origin_hf_arch == "Mistral3Model":
+                return "nemotron-3.5-embed"
+            return super().get_vocab_base_pre(tokenizer)
+
+        def set_gguf_parameters(self):
+            super().set_gguf_parameters()
+            self._try_set_pooling_type()
+
    class Mistral4Model(DeepseekV2Model):
        model_arch = gguf.MODEL_ARCH.MISTRAL4
        skip_mtp = False # model contains no MTP layers, so no need to skip
@@ -42,12 +61,20 @@ class Mistral3Model(TextModel):
    model_arch = gguf.MODEL_ARCH.MISTRAL3 # unused
    impl: TextModel

-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
+    def __init__(self, dir_model, *args, **kwargs):
+        hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        hf_arch = get_model_architecture(hparams, self.model_type)
+        if hf_arch in ("Ministral3Model", "Mistral3Model"):
+            text_config = hparams.get("text_config", hparams)
+            if text_config.get("is_causal") is not False:
+                raise NotImplementedError(f"Expected a Nemotron Embed model for {hf_arch}, but is_causal is not False")
+        super().__init__(dir_model, *args, **kwargs)
        if self.hparams.get("model_type") == "mistral4":
-            self.impl = Mistral3Model.Mistral4Model(*args, **kwargs)
+            self.impl = Mistral3Model.Mistral4Model(dir_model, *args, **kwargs)
+        elif self.hf_arch in ("Ministral3Model", "Mistral3Model"):
+            self.impl = Mistral3Model.Mistral3EmbedNemotronModel(dir_model, *args, **kwargs)
        else:
-            self.impl = Mistral3Model.Ministral3Model(*args, **kwargs)
+            self.impl = Mistral3Model.Ministral3Model(dir_model, *args, **kwargs)

    def set_vocab(self):
        self.impl.set_vocab()
@@ -1,5 +1,6 @@
 from __future__ import annotations

+import re
 from typing import Any, Callable, Iterable, Sequence, TYPE_CHECKING

 import numpy as np
@@ -367,9 +368,21 @@ class NemotronModel(TextModel):
    model_arch = gguf.MODEL_ARCH.NEMOTRON

    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-        self.gguf_writer.add_pad_token_id(0)
-        self.gguf_writer.add_unk_token_id(1)
+        if (self.dir_model / "tokenizer.model").exists():
+            self._set_vocab_sentencepiece()
+            self.gguf_writer.add_pad_token_id(0)
+            self.gguf_writer.add_unk_token_id(1)
+        else:
+            # SPM-style BPE (Metaspace + byte fallback): mark <0xXX> tokens as BYTE so they decode to raw bytes
+            tokens, toktypes, tokpre = self.get_vocab_base()
+            toktypes = [gguf.TokenType.BYTE if re.fullmatch(r"<0x[0-9A-F]{2}>", tok) else typ for tok, typ in zip(tokens, toktypes)]
+            self.gguf_writer.add_tokenizer_model("gpt2")
+            self.gguf_writer.add_tokenizer_pre(tokpre)
+            self.gguf_writer.add_token_list(tokens)
+            self.gguf_writer.add_token_types(toktypes)
+
+            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+            special_vocab.add_to_gguf(self.gguf_writer)

    def set_gguf_parameters(self):
        super().set_gguf_parameters()
@@ -383,7 +396,10 @@ class NemotronModel(TextModel):
        rot_pct = self.rope_parameters["partial_rotary_factor"]
        n_embd = self.find_hparam(["hidden_size", "n_embd"])
        n_head = self.find_hparam(["num_attention_heads", "n_head"])
-        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
+        # Use explicit head_dim when available; fallback keeps standard Nemotron working.
+        # Without this, models where n_head * head_dim != n_embd (e.g. 48 heads, head_dim=128) produce the wrong rope dimension count.
+        head_dim = self.hparams.get("head_dim") or n_embd // n_head
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * head_dim))

        # * RopeScaling for Nemotron
        factor = self.hparams.get("factor") or self.rope_parameters.get("factor")
@@ -41,6 +41,8 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
        self._ple_shards: dict[int, str] = {}
        self._ple_row_dim: int | None = None
        self._mtp_fc: dict[str, Tensor] = {}
+        # halves of the merged weights, keyed by (bid, merged tensor), see _merge_shared_input
+        self._merge_parts: dict[tuple[int, gguf.MODEL_TENSOR], dict[int, Tensor]] = {}

    @classmethod
    def filter_tensors(cls, item):
@@ -131,44 +133,79 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
        # int64 hash constants must stay exact; 1-D tensors force F32, so use KV
        if name.endswith("ple_embedding.layer_multipliers"):
            self._ple_multipliers = [int(x) for x in data_torch.tolist()]
-            return []
+            return
        if name.endswith("ple_embedding.ngram_heads_offsets"):
            self._ple_head_offsets = [int(x) for x in data_torch.tolist()]
-            return []
+            return
        if name.endswith("ple_embedding.ngram_heads_vocab_sizes"):
            self._ple_head_vocab_sizes = [int(x) for x in data_torch.tolist()]
-            return []
+            return

        if ".ngram_embedding.shard_" in name:
-            return self._place_ple_shard(data_torch, name)
+            yield from self._place_ple_shard(data_torch, name)
+            return

        # eh_proj([e ; h_s]) = fc_embedding(e) + fc_hidden(h_s) for every hc stream s
        if name.endswith((".nextn_fc_embedding.weight", ".nextn_fc_hidden.weight")):
            self._mtp_fc[name.rsplit(".", 2)[1]] = data_torch
            if len(self._mtp_fc) < 2:
-                return []
+                return
            eh = torch.cat([self._mtp_fc.pop("nextn_fc_embedding"), self._mtp_fc.pop("nextn_fc_hidden")], dim=1)
-            return [(self.format_tensor_name(gguf.MODEL_TENSOR.NEXTN_EH_PROJ, bid, ".weight"), eh)]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.NEXTN_EH_PROJ, bid, ".weight"), eh)
+            return

        # one projection feeds indexer q and k; split it, as minimax-m3 does
        if ".indexer.index_qk_proj.weight" in name:
            n_q = self.hparams["indexer_n_heads"] * self.hparams["indexer_head_dim"]
-            q = data_torch[:n_q]
-            k = data_torch[n_q:]
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_PROJ, bid, ".weight"), q),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_K_PROJ, bid, ".weight"), k),
-            ]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_PROJ, bid, ".weight"), data_torch[:n_q])
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_K_PROJ, bid, ".weight"), data_torch[n_q:])
+            return

        # Gemma zero-centred gammas the inherited norm.weight rule misses
        if name.endswith((".ple.norm_key.weight", ".ple.norm_query.weight", ".ple.norm_conv.weight",
                          ".indexer.q_layernorm.weight", ".indexer.k_layernorm.weight")):
-            return [(self.map_tensor_name(name), data_torch + 1)]
+            yield (self.map_tensor_name(name), data_torch + 1)
+            return

        if name.endswith(".ple.conv1d.weight"):
-            return [(self.map_tensor_name(name), data_torch.squeeze())]
+            yield (self.map_tensor_name(name), data_torch.squeeze())
+            return

-        return super().modify_tensors(data_torch, name, bid)
+        for new_name, data in super().modify_tensors(data_torch, name, bid):
+            yield from self._merge_shared_input(new_name, data, bid)
+
+    # matmuls that read the same input, written as one weight [a | b] so the runtime does one matmul:
+    # (merged tensor, (a, b), linear-attention layers only)
+    _SHARED_INPUT_MERGES = (
+        (gguf.MODEL_TENSOR.HC_ATTN_DOWN_INJECT, (gguf.MODEL_TENSOR.HC_ATTN_DOWN, gguf.MODEL_TENSOR.HC_ATTN_INJECT), False),
+        (gguf.MODEL_TENSOR.HC_FFN_DOWN_INJECT,  (gguf.MODEL_TENSOR.HC_FFN_DOWN,  gguf.MODEL_TENSOR.HC_FFN_INJECT),  False),
+        (gguf.MODEL_TENSOR.ATTN_QKV_GATE,       (gguf.MODEL_TENSOR.ATTN_QKV,     gguf.MODEL_TENSOR.ATTN_GATE),      True),
+        (gguf.MODEL_TENSOR.SSM_BETA_ALPHA_CAT,  (gguf.MODEL_TENSOR.SSM_BETA,     gguf.MODEL_TENSOR.SSM_ALPHA),      True),
+    )
+
+    # caches one half of a merged weight until the other half arrives, then yields [a | b]
+    def _merge_shared_input(self, new_name: str, data: Tensor, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # a pre-quantized checkpoint carries per-tensor scales, which one merged weight cannot hold
+        if bid is None or self.hparams.get("quantization_config"):
+            yield (new_name, data)
+            return
+
+        layer_types = self.hparams.get("layer_types") or []
+        is_linear = bid < len(layer_types) and layer_types[bid] == "linear_attention"
+
+        match = next(((merged, parts.index(part)) for merged, parts, linear_only in self._SHARED_INPUT_MERGES
+                      for part in parts
+                      if (is_linear or not linear_only) and self.match_model_tensor_name(new_name, part, bid)), None)
+        if match is None:
+            yield (new_name, data)
+            return
+
+        merged, idx = match
+        halves = self._merge_parts.setdefault((bid, merged), {})
+        halves[idx] = data
+        if len(halves) == 2:
+            del self._merge_parts[(bid, merged)]
+            yield (self.format_tensor_name(merged, bid), torch.cat([halves[0], halves[1]], dim=0))

    # the shards concatenate into a tensor of well over 100 GB
    # use LazyChunkedTensor here, a single shard resident at a time
@@ -214,6 +251,8 @@ class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
        super().prepare_tensors()
        if self._mtp_fc:
            raise ValueError(f"MTP projection missing its other half: {sorted(self._mtp_fc)}")
+        if self._merge_parts:
+            raise ValueError(f"merged weights missing a half: {sorted((bid, t.name) for bid, t in self._merge_parts)}")
        n_parts = self.hparams.get("split_ngram_parts", 0)
        if self._ple_shards and len(self._ple_shards) != n_parts:
            raise ValueError(
@@ -0,0 +1,114 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import torch
+
+from .base import MmprojModel, ModelBase, gguf
+from .qwen import Qwen2Model
+
+
+@ModelBase.example("sensenova/SenseNova-U1.5-8B-MoT")
+@ModelBase.register("NEOChatModel")
+class SenseNovaU1Model(Qwen2Model):
+    model_arch = gguf.MODEL_ARCH.SENSENOVA_U1
+
+    @classmethod
+    def filter_tensors(cls, item):
+        name, _ = item
+        if not name.startswith("language_model.") or "_mot_gen" in name:
+            return None
+        return super().filter_tensors(item)
+
+    def set_gguf_parameters(self):
+        if int(self.hparams.get("num_experts") or 0) > 1:
+            raise ValueError("SenseNova U1 conversion currently supports the dense understanding branch")
+        if self.rope_parameters.get("rope_type", "default") not in ("default", None):
+            raise ValueError("SenseNova U1 scaled RoPE is not supported")
+        if self.hparams.get("attention_bias", False):
+            raise ValueError("SenseNova U1 attention bias is not supported")
+        if self.hparams.get("use_sliding_window", False):
+            raise ValueError("SenseNova U1 sliding attention is not supported")
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base_spatial(self.hparams["rope_theta_hw"])
+
+    def get_tensors(self):
+        for name, gen in self.model_tensors.items():
+            if name.endswith((".q_norm_hw.weight", ".k_norm_hw.weight")):
+                continue
+            data = gen()
+            if name.endswith((".q_norm.weight", ".k_norm.weight")):
+                spatial = self.model_tensors[name.replace("_norm.weight", "_norm_hw.weight")]()
+                # Keep one weight vector; the graph normalizes its two halves independently.
+                data = torch.cat((data, spatial), dim=0)
+            yield name, data
+
+
+@ModelBase.register("NEOVisionModel")
+@ModelBase.example("sensenova/SenseNova-U1.5-8B-MoT")
+class SenseNovaU1VisionModel(MmprojModel):
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        else:
+            hparams = dict(hparams)
+
+        vision_config = dict(hparams["vision_config"])
+        # NEOVisionModel has no transformer blocks
+        # these defaults satisfy the generic mmproj metadata contract without replacing checkpoint values
+        vision_config.setdefault("image_size", 0)
+        vision_config.setdefault("intermediate_size", 0)
+        vision_config.setdefault("num_hidden_layers", 0)
+        vision_config.setdefault("num_attention_heads", 1)
+        hparams["vision_config"] = vision_config
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
+        self.preprocessor_config.setdefault("image_mean", [0.485, 0.456, 0.406])
+        self.preprocessor_config.setdefault("image_std", [0.229, 0.224, 0.225])
+
+    @classmethod
+    def filter_tensors(cls, item):
+        name, gen = item
+        if not name.startswith("vision_model."):
+            return None
+        return super().filter_tensors((name, gen))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type("sensenova_u1")
+        vision_config = self.hparams["vision_config"]
+        # These defaults match the known U1.5 checkpoint when older configs omit fields.
+        hidden_size = int(vision_config.get("hidden_size", 1024))
+        min_pixels = int(vision_config.get("min_pixels", 65536))
+        max_pixels = int(vision_config.get("max_pixels", 16777216))
+        downsample_ratio = float(vision_config.get("downsample_ratio", 0.5))
+        if downsample_ratio <= 0:
+            raise ValueError(f"SenseNova U1 downsample_ratio must be positive, got {downsample_ratio}")
+        merge_size = round(1.0 / downsample_ratio)
+        if abs((1.0 / downsample_ratio) - merge_size) > 1e-6:
+            raise ValueError(f"SenseNova U1 downsample_ratio must have an integer reciprocal, got {downsample_ratio}")
+        norm_eps = float(vision_config.get("layer_norm_eps", vision_config.get("rms_norm_eps", 1e-6)))
+        self.gguf_writer.add_vision_head_dim(hidden_size)
+        self.gguf_writer.add_vision_attention_layernorm_eps(norm_eps)
+        self.gguf_writer.add_vision_min_pixels(min_pixels)
+        self.gguf_writer.add_vision_max_pixels(max_pixels)
+        self.gguf_writer.add_vision_spatial_merge_size(merge_size)
+
+    def modify_tensors(self, data_torch, name, bid):
+        tensor_names = {
+            "vision_model.embeddings.patch_embedding.weight": "v.patch_embd.weight",
+            "vision_model.embeddings.patch_embedding.bias":   "v.patch_embd.bias",
+            "vision_model.embeddings.dense_embedding.weight": self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, 0),
+            "vision_model.embeddings.dense_embedding.bias": self.format_tensor_name(
+                gguf.MODEL_TENSOR.V_MMPROJ, 0, suffix=".bias"
+            ),
+        }
+        if name not in tensor_names:
+            raise ValueError(f"Unexpected SenseNova U1 vision tensor: {name!r}")
+        return [(tensor_names[name], data_torch)]
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, 0):
+            return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
@@ -0,0 +1,116 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+import torch
+
+from .base import ModelBase, MmprojModel, gguf, logger
+from .qwen import Qwen3Model
+
+
+def _is_soprano_checkpoint(dir_model: Path) -> bool:
+    if not (dir_model / "decoder.pth").is_file() or not (dir_model / "config.json").is_file():
+        return False
+    with open(dir_model / "config.json", encoding="utf-8") as f:
+        config = json.load(f)
+    if config.get("model_type") != "qwen3" or config.get("bos_token_id") != 3 or config.get("eos_token_id") != 3:
+        return False
+    if not (dir_model / "tokenizer.json").is_file():
+        return False
+    with open(dir_model / "tokenizer.json", encoding="utf-8") as f:
+        added = {t["id"]: t["content"] for t in json.load(f).get("added_tokens", [])}
+    return [added.get(i) for i in range(4)] == ["[UNK]", "[TEXT]", "[START]", "[STOP]"]
+
+
+@ModelBase.register_hparams_loader(_is_soprano_checkpoint)
+def _load_soprano_hparams(dir_model: Path) -> dict[str, Any]:
+    logger.info("gguf: detected Soprano checkpoint")
+    hparams = ModelBase.load_hparams(dir_model, False, guess=False)
+    # the text model stays Qwen3, the mmproj holds the decoder
+    hparams["text_config"] = {"architectures": ["SopranoTextModel"]}
+    hparams["architectures"] = ["SopranoModel"]
+    return hparams
+
+
+@ModelBase.register("SopranoTextModel")
+class SopranoTextModel(Qwen3Model):
+    model_arch = gguf.MODEL_ARCH.QWEN3
+
+    def get_vocab_base_pre(self, tokenizer) -> str:
+        return "soprano"
+
+    def set_vocab(self):
+        super().set_vocab()
+        self.gguf_writer.add_unk_token_id(0)
+
+
+@ModelBase.register("SopranoModel")
+@ModelBase.example("ekwek/Soprano-80M", "ekwek/Soprano-1.1-80M")
+class SopranoModel(MmprojModel):
+    has_vision_encoder = False
+    has_audio_encoder = False
+
+    def _load_decoder(self) -> dict[str, torch.Tensor]:
+        state = torch.load(self.dir_model / "decoder.pth", map_location="cpu", mmap=True, weights_only=True)
+        return state.get("state_dict", state)
+
+    def get_audio_config(self):
+        state = self._load_decoder()
+        n_embd, n_inp, _ = state["decoder.embed.weight"].shape
+        n_ff = state["decoder.convnext.0.pwconv1.weight"].shape[0]
+        n_layer = sum(1 for name in state if name.endswith(".pwconv1.weight"))
+        if n_inp != self.n_embd_text:
+            raise ValueError(f"Soprano decoder input dim {n_inp} does not match the backbone dim {self.n_embd_text}")
+        return {"num_hidden_layers": n_layer, "hidden_size": n_embd, "intermediate_size": n_ff}
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_clip_has_gen_audio_encoder(True)
+        self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.SOPRANO)
+        self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_gen_audio_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_gen_audio_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_gen_audio_block_count(self.hparams["num_hidden_layers"])
+        self.gguf_writer.add_gen_audio_head_count(1)
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(1e-6)
+
+    def get_tensors(self):
+        yield from self._load_decoder().items()
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if "dwconv.weight" in new_name or (name == "decoder.embed.weight" and n_dims == 3):
+            return gguf.GGMLQuantizationType.F16
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def modify_tensors(self, data_torch, name, bid):
+        if name == "head.istft.window":
+            # the ISTFT in mtmd builds its own window
+            if not torch.equal(data_torch.float(), torch.hann_window(data_torch.shape[0])):
+                raise ValueError("Soprano requires a periodic Hann window")
+            return
+        if name == "decoder.embed.weight" and data_torch.shape[-1] == 1:
+            data_torch = data_torch.squeeze(-1)
+
+        T = gguf.MODEL_TENSOR
+        if name.startswith("decoder.convnext."):
+            _, _, blk, key, *suffix = name.split(".")
+            block_map = {
+                "dwconv":  T.A_GEN_WAV_UP_DWCONV,
+                "norm":    T.A_GEN_WAV_UP_NORM,
+                "pwconv1": T.A_GEN_WAV_UP_PW1,
+                "pwconv2": T.A_GEN_WAV_UP_PW2,
+                "gamma":   T.A_GEN_WAV_UP_GAMMA,
+            }
+            yield self.format_tensor_name(block_map[key], int(blk), suffix="".join("." + s for s in suffix)), data_torch
+            return
+
+        tensor_map = {
+            "decoder.embed":            T.A_GEN_WAV_INPUT,
+            "decoder.norm":             T.A_GEN_WAV_NORM,
+            "decoder.final_layer_norm": T.A_GEN_WAV_OUTPUT_NORM,
+            "head.out":                 T.A_GEN_WAV_OUTPUT,
+        }
+        key, suffix = name.rsplit(".", 1)
+        yield self.format_tensor_name(tensor_map[key], suffix="." + suffix), data_torch
@@ -267,11 +267,15 @@ def main() -> None:
            sys.exit(1)

        if args.dspark:
-            if is_mistral_format or model_architecture != "DeepseekV4ForCausalLM":
-                logger.error("--dspark is only supported for DeepseekV4ForCausalLM")
+            if is_mistral_format or model_architecture not in ("DeepseekV4ForCausalLM", "DeepseekV41ForCausalLM"):
+                logger.error("--dspark is only supported for DeepseekV4ForCausalLM and DeepseekV41ForCausalLM")
                sys.exit(1)
-            from conversion.deepseek import DeepseekV4DSparkModel
-            model_class = DeepseekV4DSparkModel
+            if model_architecture == "DeepseekV41ForCausalLM":
+                from conversion.deepseek41 import DeepseekV41DSparkModel
+                model_class = DeepseekV41DSparkModel
+            else:
+                from conversion.deepseek import DeepseekV4DSparkModel
+                model_class = DeepseekV4DSparkModel

        if args.mtp or args.no_mtp:
            if not model_class.supports_mtp_export:
@@ -101,6 +101,7 @@ models = [
    {"name": "command-r",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
    {"name": "tiny_aya",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/tiny-aya-base", },
    {"name": "cohere2moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/North-Mini-Code-1.0", },
+    {"name": "flint",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/RecursiveSI/flint", },
    {"name": "qwen2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
    {"name": "olmo",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
    {"name": "dbrx",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
@@ -177,6 +178,7 @@ pre_computed_hashes = [
    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air", "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902"},
    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.7-Flash", "chkhsh": "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267"},
    {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
+    {"name": "domyn-small", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/domyn/Domyn-Small-v1.0", "chkhsh": "6e4d72c6e5260b90606e26a0931fa15e4b42ec6e97719d12919f2b1f93478262"},
    {"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"},
@@ -783,7 +783,7 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
 | GGML_SYCL_TARGET   | INTEL *(default)*                     | Set the SYCL target device type.            |
 | GGML_SYCL_DEVICE_ARCH | Optional                           | Set the SYCL device architecture. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
-| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
+| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. [1.](#sycl-f16-footnote) |
 | GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | GGML_SYCL_HOST_MEM_FALLBACK | ON *(default)* \|OFF *(Optional)* | Allow host memory fallback when device memory is full during quantized weight reorder. Enables inference to continue at reduced speed (reading over PCIe) instead of failing. Requires Linux kernel 6.8+. |
@@ -791,7 +791,7 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |

-1. FP32 or FP16 have different performance impact to LLM. Recommended to test them for better prompt processing performance on your models. You need to rebuild the code after change `GGML_SYCL_F16=OFF/ON`.
+1. <a id="sycl-f16-footnote"></a>FP32 or FP16 have different performance impact to LLM. Recommended to test them for better prompt processing performance on your models. You need to rebuild the code after change `GGML_SYCL_F16=OFF/ON`.

 ### Runtime

@@ -802,7 +802,8 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_DEV2DEV_MEMCPY | 0 (default), 1, 2 | Choose the method of dev2dev memory copy.<br>Value: <br>*  0: SYCL API (default), only support dGPUs.<br>* 1: L0 API -- Better performance, only support dGPUs, found to lead to abnormal crash in some case. <br>* 2: Host Forward -- Most stable method for all cases (including iGPU + dGPU*N), but with lower performance (-2% to -5%).<br>SYCL & L0 API are easy to be impacted by Intel GPU driver issue. When you meet the garbled output or crash issues in multiple GPUs case, try with this debug flag to work around or check the issue.|
 | GGML_SYCL_ENABLE_FLASH_ATTN | 1 (default) or 0| Enable Flash-Attention. It can reduce memory usage. The performance impact depends on the LLM.|
 | GGML_SYCL_ENABLE_OPT | 0 or 1 (default)| Enable optimize features for Intel GPUs. (Recommended to 0 for Intel devices older than Gen 10) |
-| GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
+| GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still in development, performance depends on model and device. [1.](#sycl-graph-footnote) |
+| GGML_SYCL_GRAPH_EVICTION_TIMEOUT | 10 (default) or positive integer | Seconds a cached SYCL graph can go unused before it is evicted. Checked on a sweep that runs every half of this timeout. |
 | GGML_SYCL_ENABLE_HOST_PINNED_MEM | 0 or 1 (default) | Enable host pinned memory to speed up copy data from host to device. When disable it, host memory will common malloc() on CPU. Disable it when use `--load-model mlock`.|
 | GGML_SYCL_HOST_PINNED_MEM_2G | 0 (default) or 1 | Limit the max memory allocation to be no more than 2GB when enable host pinned memory. USM allocations above 2 GiB take the relaxed/large-allocation path, which serializes H2D copies with compute and prevents copy/compute overlap. It will impact the startup time. Need more test. Depend on `GGML_SYCL_ENABLE_HOST_PINNED_MEM=1`.|
 | GGML_SYCL_UPLOAD_STAGING_SLOTS | 4 (default) or non-negative integer | Number of 8 MiB pinned host slots used to stage tensor uploads (model loading), so the host copy of one slot overlaps the transfer of the previous one. Set to 0 to use the old path: a malloc'd bounce buffer and a blocking copy per tensor. |
@@ -819,6 +820,9 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | 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. |
+| GGML_SYCL_ENABLE_XMX_ESIMD | 0 or 1 (default) | Use the XMX (DPAS) multi-column mat-vec kernels for reordered Q2_K to Q6_K and Q8_0 weights on Xe2 GPUs. Speeds up speculative decoding and short prompt batches. |
+| GGML_SYCL_XMX_MAX_COLS | non-negative integer, 80 (default) | Most activation columns (batch size) the `GGML_SYCL_ENABLE_XMX_ESIMD` kernels take; wider batches use the other mat-mul kernels. A larger value keeps XMX for bigger batches (as repeated launches), `0` never uses it. |
+| GGML_SYCL_XMX_GLU_MAX_COLS | 0 to 16, 16 (default) | Most columns the fused gate and up XMX kernel takes; wider batches run gate and up as separate mat-vecs. 16 is the most the kernel is built for, larger values are clamped to 16 and `0` turns the fused kernel off. |
 | GGML_SYCL_XMX_GATHER_TYPES | decimal bitmask, all bits set (default) | Weight formats that may use the XMX dequant-GEMM paths, which dequantize weights straight into the XMX tiles. This speeds up prompt processing of MoE models on GPUs with XMX units (Arc A- and B-series, Arc Pro, Data Center GPU Max), for example pp512 of Qwen3-30B-A3B UD-IQ3_XXS by about 50% on an Arc Pro B60. Bits:<br>* 1: IQ4_NL, 2: IQ3_S, 4: IQ4_XS, 8: IQ3_XXS, 16: IQ2_XXS, 32: IQ2_XS, 64: IQ2_S, 128: IQ1_S, 256: IQ1_M<br>* 512: Q8_0, 1024: Q4_K, 2048: Q5_K, 4096: Q6_K (MoE `MUL_MAT_ID` only)<br>Add values to combine them, for example `3` for IQ4_NL and IQ3_S; `0` disables the paths. A set bit does not force the path: batches of more than 64 tokens per expert or row lengths that are not a multiple of 256 (32 for IQ4_NL and Q8_0) use the library GEMM. |
 | GGML_SYCL_XMX_GATHER_SHAPES | decimal bitmask, 255 (default) | XMX `joint_matrix` combinations the paths of `GGML_SYCL_XMX_GATHER_TYPES` may use; the operand type comes from `GGML_SYCL_DYNAMIC_PRECISION` and the best supported combination is picked automatically (logged as `fg_pick_combo`). Bits:<br>* Xe2, Xe3, Xe-HPC: 1: f16 8x16x16, 2: f16 16x16x16, 4: f16 32x64x16, 8: f16 32x64x32, 32: tf32 8x16x8, 64: bf16 8x16x16<br>* Xe-HPG (Arc A770, ARL-H): 16: f16 8x8x16, 128: bf16 8x8x16<br>Clear a bit to exclude a combination, or set a single bit to force one for testing. |
 | GGML_SYCL_DYNAMIC_PRECISION | `F16` (default with `GGML_SYCL_F16=ON`), `BF16`, `TF32` or `F32` (default otherwise) | Operand type of the XMX dequant-GEMM paths (`GGML_SYCL_XMX_GATHER_TYPES`); accumulation is always f32. `F16` is the fastest, but activations above 65504 overflow. `BF16` keeps the f32 range at a 7-bit mantissa, `TF32` keeps the range and the f16 mantissa but is about 30% slower and needs Xe2, Xe3 or Xe-HPC, and `F32` turns the XMX paths off. Ops that request a higher src1 precision ([TAG_GGML_PREC]) get it regardless of this setting. |
@@ -831,6 +835,17 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |
 | GGML_SYCL_USM_SYSTEM | 0 (default) or 1 | Enable experimental support for [USM system allocations](https://github.khronos.org/SYCL_Reference/iface/usm_basic_concept.html#system-allocations) for large GPU buffers. This requires enough host memory for model weights and caches, an Intel Xe2+ GPU such as BMG or newer and supported on Linux only, with CONFIG_DRM_XE_GPUSVM enabled. |

+1. <a id="sycl-graph-footnote"></a>SYCL Graph support (`GGML_SYCL_ENABLE_GRAPH=1`) is limited to single-GPU workloads and may also depend on the model architecture and kernels in use.
+   Performance depends on how often a recorded graph can be reused, which will vary by workload.
+   For this reason, it is recommended to measure performance both with and without graphs enabled on your workloads.
+
+   To check whether graphs are supported for your workload, enable INFO logging (`-lv 4`) and look for `enabling SYCL graphs` or `disabling SYCL graphs`.
+   Support may differ between prefill and decode, so check the status for the phases that matter to you.
+   DEBUG logging (`-lv 5`) reports why graphs are disabled.
+
+   Then benchmark with your preferred tool, such as `llama-bench`.
+   For graph activity details (including reuse), set `GGML_SYCL_DEBUG=1` and look for `[SYCL-GRAPH]` messages.
+
 ## Compile-time Flags

 Pass these via `CXXFLAGS` or add a one-off `#define` to enable a flag on the spot.
@@ -2,6 +2,7 @@

 > [!WARNING]
 > **Note:** zDNN is **not** the same as ZenDNN.
+>
 > - **zDNN** (this page): IBM's Deep Neural Network acceleration library for IBM Z & LinuxONE Mainframes
 > - **ZenDNN**: AMD's deep learning library for AMD EPYC CPUs ([see ZenDNN documentation](ZenDNN.md))

@@ -46,6 +47,14 @@ The IBM zDNN backend has the following CMake options that control the behaviour
 | --------------------- | ------------- | -------------------------------------------------------------------------------------- |
 | `GGML_ZDNN_HIGH_PREC` | `0`           | Set to `1` to requantize Q8_0 weights in smaller groups, for lower loss at lower speed |

+## Performance Tunables
+
+| Flag Option             | Default Value | Description                                            |
+| ----------------------- | ------------- | ------------------------------------------------------ |
+| `--flash-attn` or `-fa` | `auto`        | Set to `on` to run attention on the NNPA (Recommended) |
+
+Note: Set `-fa on` explicitly. Flash Attention fuses each layer's attention (the KQ matrix multiplication, softmax and KQV matrix multiplication) into one `FLASH_ATTN_EXT` operation, which zDNN runs on the NNPA instead of the CPU. The speedup grows with the context length. With `auto`, llama.cpp currently turns Flash Attention off, because the operation runs on a different device than the layer. zDNN runs it when the KV cache is F16 (the default), the model uses no attention sinks, ALiBi or logit softcapping, and there is one KV cache stream (one sequence, or `--kv-unified`). Otherwise it falls back to the CPU.
+
 ## 1. Install zDNN Library

 Note: A minimum of libzdnn 1.1.2 is required for the IBM zDNN backend to link correctly.
@@ -14,7 +14,7 @@ Legend:

 | Operation | BLAS | CANN | CPU | CUDA | ET | HTP | MTL | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
 |-----------|------|------|------|------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             ADD1 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -22,12 +22,12 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        COL2IM_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                       CONV_1D_DW | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
+|                       CONV_1D_DW | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ |
 |                          CONV_2D | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -45,7 +45,7 @@ Legend:
 |                     DSV4_HC_COMB | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                     DSV4_HC_POST | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                      DSV4_HC_PRE | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              DUP | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              DUP | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -55,10 +55,10 @@ Legend:
 |                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
@@ -66,7 +66,7 @@ Legend:
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | 🟡 | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                LIGHTNING_INDEXER | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -74,12 +74,12 @@ Legend:
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                 MUL_MAT_HADAMARD | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 |
+|                 MUL_MAT_HADAMARD | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
-|                  MUL_MAT_ID_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                  MUL_MAT_ID_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                     MUL_MAT_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                     MUL_MAT_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                  MUL_MAT_ID_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                  MUL_MAT_ID_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     MUL_MAT_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     MUL_MAT_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -87,9 +87,9 @@ Legend:
 |                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                              PAD | ❌ | 🟡 | ✅ | 🟡 | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -101,7 +101,7 @@ Legend:
 |                            ROUND | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                         SET_ROWS | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -117,12 +117,12 @@ Legend:
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SUM | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                     SWIGLU_CLAMP | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     SWIGLU_CLAMP | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -8,10 +8,10 @@ extern "C" {

 #define RPC_PROTO_MAJOR_VERSION    8
 #define RPC_PROTO_MINOR_VERSION    0
-#define RPC_PROTO_PATCH_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    1

 #ifdef  __cplusplus
-static_assert(GGML_OP_COUNT == 101, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");
+static_assert(GGML_OP_COUNT == 102, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");
 #endif

 #define GGML_RPC_MAX_SERVERS       16
@@ -585,6 +585,7 @@ extern "C" {
        GGML_OP_DSV4_HC_COMB,
        GGML_OP_DSV4_HC_PRE,
        GGML_OP_DSV4_HC_POST,
+        GGML_OP_TOP_K_MASK,

        GGML_OP_UNARY,

@@ -925,6 +926,7 @@ extern "C" {
            enum   ggml_type      type);

    // dst[i0, i1, i2] = a[i0, i1, i2] + b[i0, ids[i1, i2]]
+    // an id of -1 adds nothing: dst[i0, i1, i2] = a[i0, i1, i2]
    GGML_API struct ggml_tensor * ggml_add_id(
            struct ggml_context * ctx,
            struct ggml_tensor  * a,
@@ -1476,6 +1478,16 @@ extern "C" {
            enum ggml_prec       prec,
            int                  idx);

+    // [TAG_GGML_PREC]
+    // get the accumulator / src[idx] precision set by the functions above
+    // return GGML_PREC_UNDEFINED when the op does not carry the requested precision
+    GGML_API enum ggml_prec ggml_prec_get_acc(
+            const struct ggml_tensor * a);
+
+    GGML_API enum ggml_prec ggml_prec_get_src(
+            const struct ggml_tensor * a,
+            int                        idx);
+
    // A: k columns, n rows => [ne03, ne02, n, k]
    // B: k columns, m rows  (i.e. we transpose it internally) => [ne03 * x, ne02 * y, m, k]
    // result is n columns, m rows => [ne03 * x, ne02 * y, m, n]
@@ -1497,6 +1509,7 @@ extern "C" {
            enum ggml_op_hint    hint);

    // indirect matrix multiplication
+    // an id of -1 skips the slot and zeroes the matching row of the result
    //
    // [TAG_MUL_MAT_ID_DUPLICATE_ID]
    // note: this op has undefined behaviour if a given `a` is referenced more than once for a given `b`
@@ -2732,6 +2745,26 @@ extern "C" {
            struct ggml_tensor  * post,
            struct ggml_tensor  * comb);

+    // top-k sparse attention mask: keep the mask cells of the live picks and the extra cells, set the rest to -inf
+    //
+    // mask:   [n_kv,    m1,  m2,  m3]  f16 or f32, broadcast to the rows of top_k
+    // top_k:  [n_top,   ne1, ne2, ne3] i32, picks of each row
+    // score:  [n_pick,  ne1, ne2, ne3] f32, can be NULL: a pick is live when its score is above -inf, else every pick is live
+    // blocks: [n_blk,   n_pick]        i32, can be NULL: the member cells of each pick, else a pick is a cell
+    // extra:  [n_extra, ne1, ne2, ne3] i32, can be NULL: cells of each row kept without a pick
+    // res:    [n_kv,    ne1, ne2, ne3] same type as mask
+    //
+    // res[c, i] = mask[c, i % m] when c is a cell of a live pick of row i or an extra cell of row i, else -inf
+    // picks and cells out of range are skipped
+    //
+    GGML_API struct ggml_tensor * ggml_top_k_mask(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * mask,
+            struct ggml_tensor  * top_k,
+            struct ggml_tensor  * score,
+            struct ggml_tensor  * blocks,
+            struct ggml_tensor  * extra);
+
    // custom operators

    typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata);
@@ -1046,7 +1046,8 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
            } break;
            case GGML_OP_DSV4_HC_COMB:
            case GGML_OP_DSV4_HC_PRE:
-            case GGML_OP_DSV4_HC_POST: {
+            case GGML_OP_DSV4_HC_POST:
+            case GGML_OP_TOP_K_MASK: {
                split_state = handle_generic(src_ss, /*scalar_only =*/ true);
            } break;
            case GGML_OP_UNARY: {
@@ -2226,6 +2226,7 @@ bool ggml_op_alloc_size_may_expand(enum ggml_op op) {
        case GGML_OP_CUMSUM:
        case GGML_OP_ARGSORT:
        case GGML_OP_TOP_K:
+        case GGML_OP_ARGMAX:
            return true;
        default:
            return false;
@@ -920,7 +920,7 @@ static void * ggml_backend_cann_buffer_get_base(ggml_backend_buffer_t buffer) {
 * @param dst Pointer to the destination buffer where transformed data will be
 * stored.
 */
-static void ggml_backend_cann_transform_q4_0(ggml_tensor * tensor, const void * src, void * dst) {
+static void ggml_backend_cann_transform_q4_0(const ggml_tensor * tensor, const void * src, void * dst) {
    int64_t n_elems     = ggml_nelements(tensor);
    int64_t groups      = n_elems / QK4_0;
    size_t  quant_bytes = n_elems * sizeof(uint8_t) / 2;
@@ -1014,7 +1014,7 @@ static void ggml_backend_cann_transform_back_q4_0(const ggml_tensor * tensor, vo
 * @param dst Pointer to the destination buffer where transformed data will be
 * stored.
 */
-static void ggml_backend_cann_transform_q8_0(ggml_tensor * tensor, const void * src, void * dst) {
+static void ggml_backend_cann_transform_q8_0(const ggml_tensor * tensor, const void * src, void * dst) {
    int64_t n_elems     = ggml_nelements(tensor);
    int64_t groups      = n_elems / QK8_0;
    size_t  quant_bytes = n_elems * sizeof(uint8_t);
@@ -1075,7 +1075,7 @@ static void ggml_backend_cann_transform_back_q8_0(const ggml_tensor * tensor, co
 * @param dst Pointer to the destination buffer where transformed data will be
 * stored.
 */
-static void ggml_backend_cann_transform(ggml_tensor * tensor, const void * src, void * dst) {
+static void ggml_backend_cann_transform(const ggml_tensor * tensor, const void * src, void * dst) {
    switch (tensor->type) {
        case GGML_TYPE_Q4_0:
            ggml_backend_cann_transform_q4_0(tensor, src, dst);
@@ -650,9 +650,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)

        # Fetch KleidiAI sources:
        include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.31.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.32.0")
        set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/releases/download/${KLEIDIAI_COMMIT_TAG}/kleidiai-${KLEIDIAI_COMMIT_TAG}-src.tar.gz")
-        set(KLEIDIAI_RELEASE_ARCHIVE_MD5  "930495f3aeb16b25dc18a2bde45924f4")
+        set(KLEIDIAI_RELEASE_ARCHIVE_MD5  "05f1953428ffe8e8c7f585dc6d508887")

        set(KLEIDIAI_FETCH_ARGS
            URL ${KLEIDIAI_DOWNLOAD_URL}
@@ -287,6 +287,9 @@ void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi

    sumf = vaddvq_f32(sumv);
 #else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
    ggml_vec_dot_q2_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
    return;
 #endif
@@ -1652,7 +1652,12 @@ static void ggml_compute_forward_mul_mat_id(
            for (int id = 0; id < n_ids; ++id) {
                const int32_t i02 = *(const int32_t *) ((const char *) ids->data + iid1*ids->nb[1] + id*ids->nb[0]);

-                assert(i02 >= 0 && i02 < n_as);
+                assert(i02 == -1 || (i02 >= 0 && i02 < n_as));
+
+                if (i02 == -1) {
+                    memset((char *) dst->data + id*nb1 + iid1*nb2, 0, ne0*sizeof(float));
+                    continue;
+                }

                MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = (struct mmid_row_mapping) {id, iid1};
                matrix_row_counts[i02] += 1;
@@ -2105,6 +2110,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
            {
                ggml_compute_forward_dsv4_hc_post(params, tensor);
            } break;
+        case GGML_OP_TOP_K_MASK:
+            {
+                ggml_compute_forward_top_k_mask(params, tensor);
+            } break;
        case GGML_OP_MAP_CUSTOM1:
            {
                ggml_compute_forward_map_custom1(params, tensor);
@@ -2288,6 +2297,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
        case GGML_OP_DSV4_HC_COMB:
        case GGML_OP_DSV4_HC_PRE:
        case GGML_OP_DSV4_HC_POST:
+        case GGML_OP_TOP_K_MASK:
            {
                n_tasks = n_threads;
            } break;
@@ -201,85 +201,117 @@ static struct kai_matmul_uker_config q4_k_matmul_config() {
    return config;
 }

-template<struct kai_matmul_uker_api (*GetApi)(void)>
-static size_t q4_k_get_m_step() {
-    const struct kai_matmul_uker_api api = GetApi();
-    const struct kai_matmul_uker_config config = q4_k_matmul_config();
-    return api.get_step(&config).m;
+static struct kai_matmul_uker_config default_matmul_config() {
+    return {};
 }

-template<struct kai_matmul_uker_api (*GetApi)(void)>
-static size_t q4_k_get_n_step() {
-    const struct kai_matmul_uker_api api = GetApi();
-    const struct kai_matmul_uker_config config = q4_k_matmul_config();
-    return api.get_step(&config).n;
-}
+template<
+        struct kai_matmul_uker_api (*GetApi)(void),
+        struct kai_matmul_uker_config (*GetConfig)(void),
+        size_t MR, size_t NR, size_t KR, size_t SR>
+struct matmul_uker_adapter {
+    static size_t get_m_step() {
+        const struct kai_matmul_uker_api api = GetApi();
+        const struct kai_matmul_uker_config config = GetConfig();
+        return api.get_step(&config).m;
+    }

-template<size_t Value>
-static size_t q4_k_get_constant() {
-    return Value;
-}
+    static size_t get_n_step() {
+        const struct kai_matmul_uker_api api = GetApi();
+        const struct kai_matmul_uker_config config = GetConfig();
+        return api.get_step(&config).n;
+    }

-template<struct kai_matmul_uker_api (*GetApi)(void)>
-static size_t q4_k_get_dst_offset(size_t m_idx, size_t n_idx, size_t dst_stride) {
-    const struct kai_matmul_uker_api api = GetApi();
-    const struct kai_matmul_uker_config config = q4_k_matmul_config();
-    const struct kai_matmul_uker_dst_dim_args index = { m_idx, n_idx };
-    const struct kai_matmul_uker_dst_stride_args stride = { dst_stride };
-    return api.get_dst_offset(&config, &index, &stride);
-}
+    static size_t get_mr() {
+        if constexpr (MR == 0) {
+            return get_m_step();
+        }
+        return MR;
+    }

-template<struct kai_matmul_uker_api (*GetApi)(void)>
-static size_t q4_k_get_lhs_packed_offset(size_t m_idx, size_t k, size_t bl) {
-    GGML_ASSERT(bl == KLEIDIAI_Q4_K_BLOCK_LENGTH);
+    static size_t get_nr() {
+        if constexpr (NR == 0) {
+            return get_n_step();
+        }
+        return NR;
+    }

-    const struct kai_matmul_uker_api api = GetApi();
-    const struct kai_matmul_uker_config config = q4_k_matmul_config();
-    const struct kai_matmul_uker_lhs_dim_args shape = { m_idx, k };
-    const struct kai_matmul_uker_lhs_dim_args index = { m_idx, 0 };
-    const struct kai_matmul_uker_lhs_stride_args stride = api.get_lhs_stride(&config, &shape);
-    return api.get_lhs_offset(&config, &index, &stride);
-}
+    static size_t get_kr() {
+        return KR;
+    }

-template<struct kai_matmul_uker_api (*GetApi)(void)>
-static size_t q4_k_get_rhs_packed_offset(size_t n_idx, size_t k, size_t bl) {
-    GGML_ASSERT(bl == KLEIDIAI_Q4_K_BLOCK_LENGTH);
+    static size_t get_sr() {
+        return SR;
+    }

-    const struct kai_matmul_uker_api api = GetApi();
-    const struct kai_matmul_uker_config config = q4_k_matmul_config();
-    const struct kai_matmul_uker_rhs_dim_args shape = { n_idx, k };
-    const struct kai_matmul_uker_rhs_dim_args index = { n_idx, 0 };
-    const struct kai_matmul_uker_rhs_stride_args stride = api.get_rhs_stride(&config, &shape);
-    return api.get_rhs_offset(&config, &index, &stride);
-}
+    static size_t get_dst_offset(size_t m_idx, size_t n_idx, size_t dst_stride) {
+        const struct kai_matmul_uker_api api = GetApi();
+        const struct kai_matmul_uker_config config = GetConfig();
+        const struct kai_matmul_uker_dst_dim_args index = { m_idx, n_idx };
+        const struct kai_matmul_uker_dst_stride_args stride = { dst_stride };
+        return api.get_dst_offset(&config, &index, &stride);
+    }

-template<struct kai_matmul_uker_api (*GetApi)(void)>
-static void q4_k_run_matmul(
-        size_t m, size_t n, size_t k, size_t bl,
-        const void * lhs_packed, const void * rhs_packed,
-        void * dst, size_t dst_stride_row, size_t dst_stride_col,
-        float clamp_min, float clamp_max) {
-    GGML_ASSERT(bl == KLEIDIAI_Q4_K_BLOCK_LENGTH);
-    GGML_ASSERT(dst_stride_col == sizeof(float));
+    static size_t get_lhs_packed_offset(size_t m_idx, size_t k, size_t bl) {
+        const struct kai_matmul_uker_api api = GetApi();
+        const struct kai_matmul_uker_config config = GetConfig();
+        GGML_ASSERT(bl == config.format.bl);
+        const struct kai_matmul_uker_lhs_dim_args shape = { m_idx, k };
+        const struct kai_matmul_uker_lhs_dim_args index = { m_idx, 0 };
+        const struct kai_matmul_uker_lhs_stride_args stride = api.get_lhs_stride(&config, &shape);
+        return api.get_lhs_offset(&config, &index, &stride);
+    }

-    const struct kai_matmul_uker_api api = GetApi();
-    const struct kai_matmul_uker_config config = q4_k_matmul_config();
-    const struct kai_matmul_uker_lhs_dim_args lhs_shape = { m, k };
-    const struct kai_matmul_uker_rhs_dim_args rhs_shape = { n, k };
+    static size_t get_rhs_packed_offset(size_t n_idx, size_t k, size_t bl) {
+        const struct kai_matmul_uker_api api = GetApi();
+        const struct kai_matmul_uker_config config = GetConfig();
+        GGML_ASSERT(bl == config.format.bl);
+        const struct kai_matmul_uker_rhs_dim_args shape = { n_idx, k };
+        const struct kai_matmul_uker_rhs_dim_args index = { n_idx, 0 };
+        const struct kai_matmul_uker_rhs_stride_args stride = api.get_rhs_stride(&config, &shape);
+        return api.get_rhs_offset(&config, &index, &stride);
+    }

-    struct kai_matmul_uker_args args = {};
-    args.flags = KAI_MATMUL_UKER_FLAGS_ARGS_CLAMP;
-    args.shape = { m, n, k };
-    args.operand.lhs.ptr = lhs_packed;
-    args.operand.lhs.stride = api.get_lhs_stride(&config, &lhs_shape);
-    args.operand.rhs.ptr = rhs_packed;
-    args.operand.rhs.stride = api.get_rhs_stride(&config, &rhs_shape);
-    args.operand.dst.ptr = dst;
-    args.operand.dst.stride.m = dst_stride_row;
-    args.activation.clamp.min_ptr = &clamp_min;
-    args.activation.clamp.max_ptr = &clamp_max;
-    api.run(&config, &args);
-}
+    static void run(
+            size_t m, size_t n, size_t k, size_t bl,
+            const void * lhs_packed, const void * rhs_packed,
+            void * dst, size_t dst_stride_row, size_t dst_stride_col,
+            float clamp_min, float clamp_max) {
+        const struct kai_matmul_uker_api api = GetApi();
+        const struct kai_matmul_uker_config config = GetConfig();
+        GGML_ASSERT(bl == config.format.bl);
+        GGML_ASSERT(dst_stride_col == sizeof(float));
+        const struct kai_matmul_uker_lhs_dim_args lhs_shape = { m, k };
+        const struct kai_matmul_uker_rhs_dim_args rhs_shape = { n, k };
+
+        struct kai_matmul_uker_args args = {};
+        args.flags = KAI_MATMUL_UKER_FLAGS_ARGS_CLAMP;
+        args.shape = { m, n, k };
+        args.operand.lhs.ptr = lhs_packed;
+        args.operand.lhs.stride = api.get_lhs_stride(&config, &lhs_shape);
+        args.operand.rhs.ptr = rhs_packed;
+        args.operand.rhs.stride = api.get_rhs_stride(&config, &rhs_shape);
+        args.operand.dst.ptr = dst;
+        args.operand.dst.stride.m = dst_stride_row;
+        args.activation.clamp.min_ptr = &clamp_min;
+        args.activation.clamp.max_ptr = &clamp_max;
+        api.run(&config, &args);
+    }
+};
+
+using q4_k_gemm_adapter = matmul_uker_adapter<
+    kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa,
+    q4_k_matmul_config, 0, 0, 4, 2>;
+using q4_k_gemv_adapter = matmul_uker_adapter<
+    kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot,
+    q4_k_matmul_config, 0, 0, 4, 2>;
+// GEMM and GEMV share the RHS layout only at SVE256 (NR=8, KR=8, SR=1).
+using sve_q8_gemm_adapter = matmul_uker_adapter<
+    kai_matmul_clamp_f32_qai8dxp4x8sf32_qsi8cxp4vsx8sf32bf32_16x4vs_sve_i8mm,
+    default_matmul_config, 4, 0, 8, 1>;
+using sve_q8_gemv_adapter = matmul_uker_adapter<
+    kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp8x8_1x8_sve_dot,
+    default_matmul_config, 1, 8, 8, 1>;

 static struct kai_matmul_pack_lhs_uker_config q4_k_lhs_pack_config() {
    struct kai_matmul_pack_lhs_uker_config config = {};
@@ -628,7 +660,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
        },
-        /* .required_cpu       = */ CPU_FEATURE_SVE | CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
+        /* .required_cpu       = */ CPU_FEATURE_SVE | CPU_FEATURE_SVEI8MM | CPU_FEATURE_DOTPROD,
        /* .lhs_type           = */ GGML_TYPE_F32,
        /* .rhs_type           = */ GGML_TYPE_Q4_0,
        /* .op_type            = */ GGML_TYPE_F32,
@@ -737,16 +769,16 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q4_k[] = {
    {
        /* SME2 GEMM */
        {
-            /* .get_m_step            = */ &q4_k_get_m_step<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
-            /* .get_n_step            = */ &q4_k_get_n_step<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
-            /* .get_mr                = */ &q4_k_get_m_step<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
-            /* .get_nr                = */ &q4_k_get_n_step<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
-            /* .get_kr                = */ &q4_k_get_constant<4>,
-            /* .get_sr                = */ &q4_k_get_constant<2>,
-            /* .get_dst_offset        = */ &q4_k_get_dst_offset<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
-            /* .get_lhs_offset_ex     = */ &q4_k_get_lhs_packed_offset<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
-            /* .get_rhs_packed_offset_ex = */ &q4_k_get_rhs_packed_offset<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
-            /* .run_kernel_ex         = */ &q4_k_run_matmul<kai_matmul_clamp_f32_f16p4vsx2_qai4c32p16vsx4s1s0sf16_4vsx16vs_sme2_mopa>,
+            /* .get_m_step            = */ q4_k_gemm_adapter::get_m_step,
+            /* .get_n_step            = */ q4_k_gemm_adapter::get_n_step,
+            /* .get_mr                = */ q4_k_gemm_adapter::get_mr,
+            /* .get_nr                = */ q4_k_gemm_adapter::get_nr,
+            /* .get_kr                = */ q4_k_gemm_adapter::get_kr,
+            /* .get_sr                = */ q4_k_gemm_adapter::get_sr,
+            /* .get_dst_offset        = */ q4_k_gemm_adapter::get_dst_offset,
+            /* .get_lhs_offset_ex     = */ q4_k_gemm_adapter::get_lhs_packed_offset,
+            /* .get_rhs_packed_offset_ex = */ q4_k_gemm_adapter::get_rhs_packed_offset,
+            /* .run_kernel_ex         = */ q4_k_gemm_adapter::run,
        },
        {
            /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_f16pmrx2_f32_neon,
@@ -756,16 +788,16 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q4_k[] = {
        },
        /* SME2 GEMV */
        {
-            /* .get_m_step            = */ &q4_k_get_m_step<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
-            /* .get_n_step            = */ &q4_k_get_n_step<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
-            /* .get_mr                = */ &q4_k_get_m_step<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
-            /* .get_nr                = */ &q4_k_get_n_step<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
-            /* .get_kr                = */ &q4_k_get_constant<4>,
-            /* .get_sr                = */ &q4_k_get_constant<2>,
-            /* .get_dst_offset        = */ &q4_k_get_dst_offset<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
-            /* .get_lhs_offset_ex     = */ &q4_k_get_lhs_packed_offset<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
-            /* .get_rhs_packed_offset_ex = */ &q4_k_get_rhs_packed_offset<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
-            /* .run_kernel_ex         = */ &q4_k_run_matmul<kai_matmul_clamp_f32_qsi8d32p1x4_qai4c32p16vsx4s1s0sf16_1x16vs_sme2_dot>,
+            /* .get_m_step            = */ q4_k_gemv_adapter::get_m_step,
+            /* .get_n_step            = */ q4_k_gemv_adapter::get_n_step,
+            /* .get_mr                = */ q4_k_gemv_adapter::get_mr,
+            /* .get_nr                = */ q4_k_gemv_adapter::get_nr,
+            /* .get_kr                = */ q4_k_gemv_adapter::get_kr,
+            /* .get_sr                = */ q4_k_gemv_adapter::get_sr,
+            /* .get_dst_offset        = */ q4_k_gemv_adapter::get_dst_offset,
+            /* .get_lhs_offset_ex     = */ q4_k_gemv_adapter::get_lhs_packed_offset,
+            /* .get_rhs_packed_offset_ex = */ q4_k_gemv_adapter::get_rhs_packed_offset,
+            /* .run_kernel_ex         = */ q4_k_gemv_adapter::run,
        },
        {
            /* .get_offset            = */ q4_k_lhs_get_offset,
@@ -883,6 +915,54 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
        /* .rhs_type           = */ GGML_TYPE_Q8_0,
        /* .op_type            = */ GGML_TYPE_F32,
    },
+    {
+        /* SVE I8MM GEMM */
+        {
+            /* .get_m_step            = */ sve_q8_gemm_adapter::get_m_step,
+            /* .get_n_step            = */ sve_q8_gemm_adapter::get_n_step,
+            /* .get_mr                = */ sve_q8_gemm_adapter::get_mr,
+            /* .get_nr                = */ sve_q8_gemm_adapter::get_nr,
+            /* .get_kr                = */ sve_q8_gemm_adapter::get_kr,
+            /* .get_sr                = */ sve_q8_gemm_adapter::get_sr,
+            /* .get_dst_offset        = */ sve_q8_gemm_adapter::get_dst_offset,
+            /* .get_lhs_offset_ex     = */ sve_q8_gemm_adapter::get_lhs_packed_offset,
+            /* .get_rhs_packed_offset_ex = */ sve_q8_gemm_adapter::get_rhs_packed_offset,
+            /* .run_kernel_ex         = */ sve_q8_gemm_adapter::run,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* SVE DOT GEMV */
+        {
+            /* .get_m_step            = */ sve_q8_gemv_adapter::get_m_step,
+            /* .get_n_step            = */ sve_q8_gemv_adapter::get_n_step,
+            /* .get_mr                = */ sve_q8_gemv_adapter::get_mr,
+            /* .get_nr                = */ sve_q8_gemv_adapter::get_nr,
+            /* .get_kr                = */ sve_q8_gemv_adapter::get_kr,
+            /* .get_sr                = */ sve_q8_gemv_adapter::get_sr,
+            /* .get_dst_offset        = */ sve_q8_gemv_adapter::get_dst_offset,
+            /* .get_lhs_offset_ex     = */ sve_q8_gemv_adapter::get_lhs_packed_offset,
+            /* .get_rhs_packed_offset_ex = */ sve_q8_gemv_adapter::get_rhs_packed_offset,
+            /* .run_kernel_ex         = */ sve_q8_gemv_adapter::run,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qai8dxp_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_quant_pack_qai8dxp_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+            /* .pack_func_ex          = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_SVE | CPU_FEATURE_SVEI8MM | CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q8_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
    {
        /* I8MM GEMM */
        {
@@ -13,7 +13,8 @@ enum cpu_feature {
    CPU_FEATURE_SVE     = 4,
    CPU_FEATURE_SME     = 8,
    CPU_FEATURE_SME2    = 16,
-    CPU_FEATURE_FP16    = 32
+    CPU_FEATURE_FP16    = 32,
+    CPU_FEATURE_SVEI8MM = 64
 };

 inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
@@ -319,8 +319,10 @@ static void init_kleidiai_context(void) {

        size_t detected_smcus = 0;

+        // The fixed-width SVE kernels require 32-byte vectors, including the Q8 shared RHS layout.
        ctx.features  = (runtime_feat.has_dotprod  ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
                        (runtime_feat.has_i8mm     ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
+                        (runtime_feat.has_svei8mm  ? CPU_FEATURE_SVEI8MM : CPU_FEATURE_NONE) |
                        (runtime_feat.has_fp16     ? CPU_FEATURE_FP16    : CPU_FEATURE_NONE) |
                        (runtime_feat.sve_cnt == QK8_0 ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);

@@ -1616,6 +1618,11 @@ public:
            const size_t kr = kernel->get_kr();
            const size_t sr = kernel->get_sr();
            const ggml_type rhs_type = kernels->rhs_type;
+            if (rhs_type == GGML_TYPE_Q8_0 && (kernels->required_cpu & CPU_FEATURE_SVE)) {
+                GGML_ASSERT(nr == kernels->gemv.get_nr());
+                GGML_ASSERT(kr == kernels->gemv.get_kr());
+                GGML_ASSERT(sr == kernels->gemv.get_sr());
+            }
            const size_t block_len = rhs_type == GGML_TYPE_Q8_0 ? QK8_0 :
                                     rhs_type == GGML_TYPE_Q4_0 ? QK4_0 :
                                     rhs_type == GGML_TYPE_Q4_K ? GGML_KLEIDIAI_Q4_K_BL :
@@ -743,12 +743,17 @@ static void ggml_compute_forward_add_id_f32(
        // src1 indices
        const int i11 = *(int32_t *) ((char *) src2->data + i1*nb20 + i2*nb21);

-        GGML_ASSERT(i11 >= 0 && i11 < ne11);
+        GGML_ASSERT(i11 == -1 || (i11 >= 0 && i11 < ne11));

-        ggml_vec_add_f32(ne0,
-                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ),
-                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
-                (float *) ((char *) src1->data + i11*nb11));
+        float       * dst_row  = (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
+        const float * src0_row = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+
+        if (i11 == -1) {
+            memcpy(dst_row, src0_row, ne0*sizeof(float));
+            continue;
+        }
+
+        ggml_vec_add_f32(ne0, dst_row, src0_row, (float *) ((char *) src1->data + i11*nb11));
    }
 }

@@ -1909,213 +1914,87 @@ void ggml_compute_forward_repeat_back(
 }

 // ggml_compute_forward_concat
-static void ggml_compute_forward_concat_any(
-    const ggml_compute_params * params,
-    ggml_tensor * dst) {

-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
+// rows of dst copied together, so that a strided source is read along neighboring rows
+static constexpr int64_t GGML_CONCAT_TILE = 32;

-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int32_t dim = ggml_get_op_params_i32(dst, 0);
-
-    GGML_ASSERT(dim >= 0 && dim < 4);
-    GGML_ASSERT(ggml_is_contiguous_rows(src0));
-    GGML_ASSERT(ggml_is_contiguous_rows(src1));
-
-    int64_t o[4] = {0, 0, 0, 0};
-
-    if (dim == 0) {
-        GGML_ASSERT(src0->ne[0] % ggml_blck_size(src0->type) == 0);
-        GGML_ASSERT(src1->ne[0] % ggml_blck_size(src1->type) == 0);
-
-        o[dim] = src0->ne[dim]/ggml_blck_size(src0->type);
-    } else {
-        o[dim] = src0->ne[dim];
-    }
-
-    // Region 1: copy rows from src0
-    for (int i3 = 0; i3 < ne03; i3++) {
-        for (int i2 = ith; i2 < ne02; i2 += nth) {
-            for (int i1 = 0; i1 < ne01; i1++) {
-                const char * x = (const char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03;
-                      char * y = (      char *) dst->data  + i1*nb1  + i2*nb2  + i3*nb3;
-                memcpy(y, x, ggml_row_size(src0->type, ne00));
-            }
-        }
-    }
-
-    // Region 2: copy rows from src1, offset into dst by o[]
-    for (int i3 = 0; i3 < ne13; i3++) {
-        for (int i2 = ith; i2 < ne12; i2 += nth) {
-            for (int i1 = 0; i1 < ne11; i1++) {
-                const char * x = (const char *) src1->data + i1*nb11         + i2*nb12         + i3*nb13;
-                      char * y = (      char *)  dst->data + (i1 + o[1])*nb1 + (i2 + o[2])*nb2 + (i3 + o[3])*nb3 + o[0]*nb0;
-                memcpy(y, x, ggml_row_size(src1->type, ne10));
-            }
+// copy n_rows rows of n elements of es bytes from a source with element stride nb0, element by element
+template <size_t es>
+static void ggml_concat_copy_strided(char * y, size_t nby, const char * x, size_t nbx, size_t nb0, int64_t n, int64_t n_rows) {
+    for (int64_t i0 = 0; i0 < n; i0++) {
+        for (int64_t r = 0; r < n_rows; r++) {
+            memcpy(y + r*nby + i0*es, x + r*nbx + i0*nb0, es);
        }
    }
 }

-static void ggml_compute_forward_concat_i8(
-    const ggml_compute_params * params,
-    ggml_tensor * dst) {
+// copy n_rows rows of n elements of src into dst, rows are nbx bytes apart in src and nby bytes apart in dst
+static void ggml_concat_copy_rows(char * y, size_t nby, const ggml_tensor * src, const char * x, size_t nbx, int64_t n, int64_t n_rows) {
+    const size_t es = ggml_type_size(src->type);

-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    GGML_ASSERT(ggml_type_size(src0->type) == sizeof(int8_t));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int32_t dim = ggml_get_op_params_i32(dst, 0);
-
-    GGML_ASSERT(dim >= 0 && dim < 4);
-
-    int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = src0->ne[dim];
-
-    const int8_t * x;
-
-    // TODO: smarter multi-theading
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ith; i2 < ne2; i2 += nth) {
-            for (int i1 = 0; i1 < ne1; i1++) {
-                for (int i0 = 0; i0 < ne0; i0++) {
-                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-                        x = (const int8_t *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
-                    } else {
-                        x = (const int8_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
-                    }
-
-                    int8_t * y = (int8_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
-
-                    *y = *x;
-                }
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_concat_f16(
-    const ggml_compute_params * params,
-    ggml_tensor * dst) {
-
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    GGML_ASSERT(ggml_type_size(src0->type) == sizeof(ggml_fp16_t));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int32_t dim = ggml_get_op_params_i32(dst, 0);
-
-    GGML_ASSERT(dim >= 0 && dim < 4);
-
-    int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = src0->ne[dim];
-
-    const ggml_fp16_t * x;
-
-    // TODO: smarter multi-theading
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ith; i2 < ne2; i2 += nth) {
-            for (int i1 = 0; i1 < ne1; i1++) {
-                for (int i0 = 0; i0 < ne0; i0++) {
-                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-                        x = (const ggml_fp16_t *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
-                    } else {
-                        x = (const ggml_fp16_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
-                    }
-
-                    ggml_fp16_t * y = (ggml_fp16_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
-
-                    *y = *x;
-                }
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_concat_f32(
-    const ggml_compute_params * params,
-    ggml_tensor * dst) {
-
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    GGML_ASSERT(ggml_type_size(src0->type) == sizeof(float));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int32_t dim = ggml_get_op_params_i32(dst, 0);
-
-    GGML_ASSERT(dim >= 0 && dim < 4);
-
-    int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = src0->ne[dim];
-
-    const float * x;
-
-    // TODO: smarter multi-theading
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ith; i2 < ne2; i2 += nth) {
-            for (int i1 = 0; i1 < ne1; i1++) {
-                for (int i0 = 0; i0 < ne0; i0++) {
-                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-                        x = (const float *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
-                    } else {
-                        x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
-                    }
-
-                    float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
-
-                    *y = *x;
-                }
-            }
+    if (src->nb[0] == es) {
+        const size_t row_size = ggml_row_size(src->type, n);
+        for (int64_t r = 0; r < n_rows; r++) {
+            memcpy(y + r*nby, x + r*nbx, row_size);
        }
+        return;
+    }
+
+    switch (es) {
+        case 1: ggml_concat_copy_strided<1>(y, nby, x, nbx, src->nb[0], n, n_rows); break;
+        case 2: ggml_concat_copy_strided<2>(y, nby, x, nbx, src->nb[0], n, n_rows); break;
+        case 4: ggml_concat_copy_strided<4>(y, nby, x, nbx, src->nb[0], n, n_rows); break;
+        case 8: ggml_concat_copy_strided<8>(y, nby, x, nbx, src->nb[0], n, n_rows); break;
+        default: GGML_ABORT("unsupported element size");
    }
 }

+// each dst row comes from src0 or src1, or for dim 0 from the row of src0 followed by the row of src1
 void ggml_compute_forward_concat(
    const ggml_compute_params * params,
    ggml_tensor * dst) {

    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];

-    switch (src0->type) {
-        case GGML_TYPE_F16:
-        case GGML_TYPE_BF16:
-        case GGML_TYPE_I16:
-            {
-                ggml_compute_forward_concat_f16(params, dst);
-            } break;
-        case GGML_TYPE_I8:
-            {
-                ggml_compute_forward_concat_i8(params, dst);
-            } break;
-        case GGML_TYPE_F32:
-        case GGML_TYPE_I32:
-            {
-                ggml_compute_forward_concat_f32(params, dst);
-            } break;
-        default:
-            {
-                ggml_compute_forward_concat_any(params, dst);
-            }
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
+    GGML_ASSERT(nb0 == ggml_type_size(dst->type));
+    GGML_ASSERT(!ggml_is_quantized(dst->type) || (nb00 == nb0 && nb10 == nb0));
+    GGML_ASSERT(dim != 0 || (ne00 % ggml_blck_size(dst->type) == 0 && ne10 % ggml_blck_size(dst->type) == 0));
+
+    const size_t y1 = dim == 0 ? ggml_row_size(dst->type, ne00) : 0;
+
+    const int64_t nr = ne1*ne2*ne3;
+    const int64_t dr = (nr + params->nth - 1)/params->nth;
+
+    const int64_t ir0 = dr*params->ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t ir = ir0; ir < ir1; ) {
+        int64_t i[4] = { 0, ir % ne1, (ir/ne1) % ne2, ir/(ne1*ne2) };
+
+        // a tile of consecutive rows sharing i2, i3 and their source
+        int64_t nt = MIN(MIN(ir1 - ir, ne1 - i[1]), GGML_CONCAT_TILE);
+        if (dim == 1 && i[1] < ne01) {
+            nt = MIN(nt, ne01 - i[1]);
+        }
+
+        char * y = (char *) dst->data + i[1]*nb1 + i[2]*nb2 + i[3]*nb3;
+
+        if (dim == 0 || i[dim] < src0->ne[dim]) {
+            ggml_concat_copy_rows(y, nb1, src0, (const char *) src0->data + i[1]*nb01 + i[2]*nb02 + i[3]*nb03, nb01, ne00, nt);
+        }
+
+        if (dim == 0 || i[dim] >= src0->ne[dim]) {
+            i[dim] -= src0->ne[dim]; // i[0] does not address rows
+            ggml_concat_copy_rows(y + y1, nb1, src1, (const char *) src1->data + i[1]*nb11 + i[2]*nb12 + i[3]*nb13, nb11, ne10, nt);
+        }
+
+        ir += nt;
    }
 }

@@ -11736,6 +11615,92 @@ void ggml_compute_forward_dsv4_hc_post(
    }
 }

+// ggml_compute_forward_top_k_mask
+
+void ggml_compute_forward_top_k_mask(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+    const ggml_tensor * mask   = dst->src[0];
+    const ggml_tensor * top_k  = dst->src[1];
+    const ggml_tensor * score  = dst->src[2];
+    const ggml_tensor * blocks = dst->src[3];
+    const ggml_tensor * extra  = dst->src[4];
+
+    GGML_ASSERT(dst->type == mask->type);
+    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
+
+    const int64_t n_kv    = mask->ne[0];
+    const int64_t ne1     = dst->ne[1];
+    const int64_t ne2     = dst->ne[2];
+    const int64_t n_rows  = ggml_nrows(dst);
+    const int64_t n_top   = top_k->ne[0];
+    const int64_t n_blk   = blocks ? blocks->ne[0] : 1;
+    const int64_t n_pick  = blocks ? blocks->ne[1] : score ? score->ne[0] : n_kv;
+    const int64_t n_extra = extra ? extra->ne[0] : 0;
+
+    const size_t ts = ggml_type_size(dst->type);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    for (int64_t ir = ith; ir < n_rows; ir += nth) {
+        const int64_t i1 = ir % ne1;
+        const int64_t i2 = (ir / ne1) % ne2;
+        const int64_t i3 = ir / (ne1*ne2);
+
+        const char * mask_row = (const char *) mask->data
+            + (i1 % mask->ne[1])*mask->nb[1] + (i2 % mask->ne[2])*mask->nb[2] + (i3 % mask->ne[3])*mask->nb[3];
+        char * dst_row = (char *) dst->data + i1*dst->nb[1] + i2*dst->nb[2] + i3*dst->nb[3];
+
+        if (dst->type == GGML_TYPE_F16) {
+            const ggml_fp16_t ninf = GGML_CPU_FP32_TO_FP16(-INFINITY);
+            for (int64_t c = 0; c < n_kv; ++c) {
+                ((ggml_fp16_t *) dst_row)[c] = ninf;
+            }
+        } else {
+            for (int64_t c = 0; c < n_kv; ++c) {
+                ((float *) dst_row)[c] = -INFINITY;
+            }
+        }
+
+        auto keep = [&](int32_t c) {
+            if (c >= 0 && c < n_kv) {
+                memcpy(dst_row + c*ts, mask_row + c*ts, ts);
+            }
+        };
+
+        const char * top_k_row = (const char *) top_k->data + i1*top_k->nb[1] + i2*top_k->nb[2] + i3*top_k->nb[3];
+
+        for (int64_t j = 0; j < n_top; ++j) {
+            const int32_t p = *(const int32_t *) (top_k_row + j*top_k->nb[0]);
+            if (p < 0 || p >= n_pick) {
+                continue;
+            }
+            if (score) {
+                const float s = *(const float *) ((const char *) score->data
+                    + p*score->nb[0] + i1*score->nb[1] + i2*score->nb[2] + i3*score->nb[3]);
+                if (!(s > -INFINITY)) {
+                    continue;
+                }
+            }
+            if (!blocks) {
+                keep(p);
+                continue;
+            }
+            for (int64_t k = 0; k < n_blk; ++k) {
+                keep(*(const int32_t *) ((const char *) blocks->data + k*blocks->nb[0] + p*blocks->nb[1]));
+            }
+        }
+
+        if (extra) {
+            const char * extra_row = (const char *) extra->data + i1*extra->nb[1] + i2*extra->nb[2] + i3*extra->nb[3];
+            for (int64_t j = 0; j < n_extra; ++j) {
+                keep(*(const int32_t *) (extra_row + j*extra->nb[0]));
+            }
+        }
+    }
+}
+
 // ggml_compute_forward_rwkv_wkv7

 static void ggml_compute_forward_rwkv_wkv7_f32(
@@ -108,6 +108,7 @@ void ggml_compute_forward_lightning_indexer(const struct ggml_compute_params * p
 void ggml_compute_forward_dsv4_hc_comb(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_dsv4_hc_pre(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_dsv4_hc_post(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_top_k_mask(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -4855,7 +4855,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
                    const int32_t i02 =
                        *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]);

-                    GGML_ASSERT(i02 >= 0 && i02 < n_as);
+                    GGML_ASSERT(i02 == -1 || (i02 >= 0 && i02 < n_as));
+
+                    if (i02 == -1) {
+                        memset((char *) dst->data + id*nb1 + iid1*nb2, 0, ne0*sizeof(float));
+                        continue;
+                    }

                    MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 };
                    matrix_row_counts[i02] += 1;
@@ -42,6 +42,7 @@ extern "C" {
    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)

    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)

    static inline float neon_compute_fp16_to_fp32(ggml_fp16_t h) {
        __fp16 tmp;
@@ -150,9 +151,15 @@ extern float ggml_table_f32_ue4m3[1 << 8];
 #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
 #endif

-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
+// a branch that defines a GGML_CPU_COMPUTE_* macro must also select it, or the fallback takes over silently
+#if defined(GGML_CPU_COMPUTE_FP16_TO_FP32) && !defined(GGML_CPU_FP16_TO_FP32) && !defined(__F16C__) // see: https://github.com/ggml-org/llama.cpp/issues/30342
+#error "a branch defined GGML_CPU_COMPUTE_FP16_TO_FP32 without selecting it as GGML_CPU_FP16_TO_FP32"
+#endif
+#if defined(GGML_CPU_COMPUTE_FP32_TO_FP16) && !defined(GGML_CPU_FP32_TO_FP16)
+#error "a branch defined GGML_CPU_COMPUTE_FP32_TO_FP16 without selecting it as GGML_CPU_FP32_TO_FP16"
+#endif
+
+// fall back to the lookup table / the portable software conversion for whichever direction no branch selected
 #if !defined(GGML_CPU_FP16_TO_FP32)
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
    uint16_t s;
@@ -682,7 +682,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_
                    const int32_t i02 =
                        *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]);

-                    GGML_ASSERT(i02 >= 0 && i02 < n_as);
+                    GGML_ASSERT(i02 == -1 || (i02 >= 0 && i02 < n_as));
+
+                    if (i02 == -1) {
+                        memset((char *) dst->data + id * nb1 + iid1 * nb2, 0, ne0 * sizeof(float));
+                        continue;
+                    }

                    MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 };
                    matrix_row_counts[i02] += 1;
@@ -78,7 +78,7 @@ static inline float op_expm1(float x) {
 }

 static inline float op_softplus(float x) {
-    return (x > 20.0f) ? x : logf(1.0f + expf(x));
+    return (x > 20.0f) ? x : log1pf(expf(x));
 }

 static inline float op_floor(float x) {
@@ -460,6 +460,92 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
    }
 }

+#ifndef GGML_GELU_QUICK_FP16
+// computed in f32 without the fp16 table, which loses precision on f32 input
+void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
+    int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i, ggml_v_gelu_quick(_mm512_loadu_ps(x + i)));
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i, ggml_v_gelu_quick(_mm256_loadu_ps(x + i)));
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        _mm_storeu_ps(y + i, ggml_v_gelu_quick(_mm_loadu_ps(x + i)));
+    }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+    const int vlen = svcntw();
+    for (; i < n; i += vlen) {
+        const svbool_t pg = svwhilelt_b32_s32(i, n);
+        svst1_f32(pg, y + i, ggml_v_gelu_quick(pg, svld1_f32(pg, x + i)));
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        vst1q_f32(y + i, ggml_v_gelu_quick(vld1q_f32(x + i)));
+    }
+#elif defined(__VXE__) || defined(__VXE2__)
+    for (; i + 3 < n; i += 4) {
+        vec_xst(ggml_v_gelu_quick(vec_xl(0, x + i)), 0, y + i);
+    }
+#elif defined(__riscv_v_intrinsic)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+        vfloat32m2_t vy = ggml_v_gelu_quick_m2(vx, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = ggml_gelu_quick_f32(x[i]);
+    }
+}
+
+void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_gelu_quick(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_gelu_quick(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        _mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_gelu_quick(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
+    }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+    const int vlen = svcntw();
+    for (; i < n; i += vlen) {
+        const svbool_t pg = svwhilelt_b32_s32(i, n);
+        svst1_f32(pg, y + i, svmul_f32_x(pg, ggml_v_gelu_quick(pg, svld1_f32(pg, x + i)), svld1_f32(pg, g + i)));
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        vst1q_f32(y + i, vmulq_f32(ggml_v_gelu_quick(vld1q_f32(x + i)), vld1q_f32(g + i)));
+    }
+#elif defined(__VXE__) || defined(__VXE2__)
+    for (; i + 3 < n; i += 4) {
+        vec_xst(vec_mul(ggml_v_gelu_quick(vec_xl(0, x + i)), vec_xl(0, g + i)), 0, y + i);
+    }
+#elif defined(__riscv_v_intrinsic)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e32m2(n - i);
+        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+        vfloat32m2_t vg = __riscv_vle32_v_f32m2(&g[i], vl);
+        vfloat32m2_t vy = __riscv_vfmul_vv_f32m2(ggml_v_gelu_quick_m2(vx, vl), vg, vl);
+        __riscv_vse32_v_f32m2(&y[i], vy, vl);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
+    }
+}
+#endif
+
 ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean) {
    int i = 0;
    ggml_float sum = 0;
@@ -44,7 +44,6 @@ inline static ggml_float ggml_sve_sum_f32x2(svfloat32_t sum_lo, svfloat32_t sum_
 #endif

 #define GGML_GELU_FP16
-#define GGML_GELU_QUICK_FP16

 #define GGML_SOFT_MAX_UNROLL 4
 #define GGML_VEC_DOT_UNROLL  2
@@ -1054,11 +1053,7 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
    }
 }
 #else
-inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
-    for (int i = 0; i < n; ++i) {
-        y[i] = ggml_gelu_quick_f32(x[i]);
-    }
-}
+void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x);
 #endif

 // Sigmoid Linear Unit (SiLU) function
@@ -1143,6 +1138,13 @@ inline static svfloat32_t ggml_v_silu(svbool_t pg, svfloat32_t x) {
    return svdiv_f32_x(pg, x, one_plus_exp_neg_x);
 }

+// gelu_quick(x) = x / (1 + exp(-1.702 x)); x stays the numerator so it cannot overflow
+inline static svfloat32_t ggml_v_gelu_quick(svbool_t pg, svfloat32_t x) {
+    const svfloat32_t one = svdup_n_f32_x(pg, 1.0f);
+    const svfloat32_t exp_cx = ggml_v_expf(pg, svmul_n_f32_x(pg, x, GELU_QUICK_COEF));
+    return svdiv_f32_x(pg, x, svadd_f32_x(pg, one, exp_cx));
+}
+
 #elif defined(__ARM_NEON) && defined(__aarch64__)

 // adapted from arm limited optimized routine
@@ -1182,6 +1184,13 @@ inline static float32x4_t ggml_v_silu(float32x4_t x) {
    return vdivq_f32(x, one_plus_exp_neg_x);
 }

+// gelu_quick(x) = x / (1 + exp(-1.702 x)); x stays the numerator so it cannot overflow
+inline static float32x4_t ggml_v_gelu_quick(float32x4_t x) {
+    const float32x4_t one = vdupq_n_f32(1.0f);
+    const float32x4_t exp_cx = ggml_v_expf(vmulq_n_f32(x, GELU_QUICK_COEF));
+    return vdivq_f32(x, vaddq_f32(one, exp_cx));
+}
+
 #elif defined(__AVX512F__) && defined(__AVX512DQ__)

 // adapted from arm limited optimized routine
@@ -1225,6 +1234,13 @@ inline static __m512 ggml_v_silu(__m512 x) {
    return _mm512_div_ps(x, one_plus_exp_neg_x);
 }

+// gelu_quick(x) = x / (1 + exp(-1.702 x)); x stays the numerator so it cannot overflow
+inline static __m512 ggml_v_gelu_quick(__m512 x) {
+    const __m512 one = _mm512_set1_ps(1);
+    const __m512 exp_cx = ggml_v_expf(_mm512_mul_ps(x, _mm512_set1_ps(GELU_QUICK_COEF)));
+    return _mm512_div_ps(x, _mm512_add_ps(one, exp_cx));
+}
+
 #elif defined(__AVX2__) && defined(__FMA__)

 // adapted from arm limited optimized routine
@@ -1280,6 +1296,13 @@ inline static __m256 ggml_v_silu(__m256 x) {
    return _mm256_div_ps(x, one_plus_exp_neg_x);
 }

+// gelu_quick(x) = x / (1 + exp(-1.702 x)); x stays the numerator so it cannot overflow
+inline static __m256 ggml_v_gelu_quick(__m256 x) {
+    const __m256 one = _mm256_set1_ps(1);
+    const __m256 exp_cx = ggml_v_expf(_mm256_mul_ps(x, _mm256_set1_ps(GELU_QUICK_COEF)));
+    return _mm256_div_ps(x, _mm256_add_ps(one, exp_cx));
+}
+
 #elif defined(__SSE2__) // __AVX2__ / __ARM_NEON

 #if defined(__FMA__)
@@ -1334,6 +1357,13 @@ inline static __m128 ggml_v_silu(__m128 x) {
    return _mm_div_ps(x, one_plus_exp_neg_x);
 }

+// gelu_quick(x) = x / (1 + exp(-1.702 x)); x stays the numerator so it cannot overflow
+inline static __m128 ggml_v_gelu_quick(__m128 x) {
+    const __m128 one = _mm_set1_ps(1);
+    const __m128 exp_cx = ggml_v_expf(_mm_mul_ps(x, _mm_set1_ps(GELU_QUICK_COEF)));
+    return _mm_div_ps(x, _mm_add_ps(one, exp_cx));
+}
+
 #elif defined(__VXE__) || defined(__VXE2__)

 // adapted from arm limited optimized routine
@@ -1374,6 +1404,13 @@ inline static float32x4_t ggml_v_silu(float32x4_t v_x) {
    return vec_div(v_x, v_one_plus_exp_neg_x);
 }

+// gelu_quick(x) = x / (1 + exp(-1.702 x)); x stays the numerator so it cannot overflow
+inline static float32x4_t ggml_v_gelu_quick(float32x4_t x) {
+    const float32x4_t one = vec_splats(1.0f);
+    const float32x4_t exp_cx = ggml_v_expf(vec_mul(x, vec_splats(GELU_QUICK_COEF)));
+    return vec_div(x, vec_add(one, exp_cx));
+}
+
 #elif defined(__riscv_v_intrinsic)

 // adapted from arm limited optimized routine
@@ -1426,6 +1463,12 @@ inline static vfloat32m2_t ggml_v_silu_m2(vfloat32m2_t x, int vl) {
    return __riscv_vfdiv_vv_f32m2(x, one_plus_exp_neg_x, vl);
 }

+// gelu_quick(x) = x / (1 + exp(-1.702 x)); x stays the numerator so it cannot overflow
+inline static vfloat32m2_t ggml_v_gelu_quick_m2(vfloat32m2_t x, int vl) {
+    const vfloat32m2_t exp_cx = ggml_v_expf_m2(__riscv_vfmul_vf_f32m2(x, GELU_QUICK_COEF, vl), vl);
+    return __riscv_vfdiv_vv_f32m2(x, __riscv_vfadd_vf_f32m2(exp_cx, 1.0f, vl), vl);
+}
+
 #endif // __ARM_NEON / __AVX2__ / __SSE2__ / __VXE__ / __riscv_v_intrinsic

 inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
@@ -1571,11 +1614,7 @@ inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float
    }
 }
 #else
-inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
-    for (int i = 0; i < n; ++i) {
-        y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
-    }
-}
+void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g);
 #endif

 inline static void ggml_vec_geglu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
@@ -1264,15 +1264,12 @@ struct ggml_tensor_extra_gpu {
 struct ggml_cuda_graph {
 #ifdef USE_CUDA_GRAPH
    ~ggml_cuda_graph() {
-        if (instance != nullptr) {
+        for (cudaGraphExec_t instance : instances) {
            CUDA_CHECK(cudaGraphExecDestroy(instance));
        }
-        if (graph != nullptr) {
-            CUDA_CHECK(cudaGraphDestroy(graph));
-        }
    }
-    cudaGraph_t graph = nullptr;
-    cudaGraphExec_t instance = nullptr;
+    // the ggml graph is captured as a series of CUDA graphs, so the GPU can run one chunk while the next is enqueued
+    std::vector<cudaGraphExec_t> instances;
    size_t num_nodes = 0;
    std::vector<cudaGraphNode_t> nodes;
    bool disable_due_to_gpu_arch = false;
@@ -1501,7 +1498,7 @@ struct ggml_backend_cuda_context {
    // Check if any CUDA graph has an instance for this context
    bool any_cuda_graph_has_instance() const {
        for (const auto & [key, graph] : cuda_graphs) {
-            if (graph && graph->instance != nullptr) {
+            if (graph && !graph->instances.empty()) {
                return true;
            }
        }
@@ -1551,11 +1548,11 @@ struct ggml_backend_cuda_context {
    // pool
    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];

-    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device, int stream_no);
+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(ggml_backend_cuda_context & ctx, int device, int stream_no);

    ggml_cuda_pool & pool(int device) {
        if (pools[device][curr_stream_no] == nullptr) {
-            pools[device][curr_stream_no] = new_pool_for_device(device, curr_stream_no);
+            pools[device][curr_stream_no] = new_pool_for_device(*this, device, curr_stream_no);
        }
        return *pools[device][curr_stream_no];
    }
@@ -70,6 +70,7 @@
 #include "ggml-cuda/cumsum.cuh"
 #include "ggml-cuda/fill.cuh"
 #include "ggml-cuda/lightning-indexer.cuh"
+#include "ggml-cuda/top-k-mask.cuh"
 #include "ggml.h"

 #include <algorithm>
@@ -213,52 +214,36 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
 }

 #if defined(GGML_USE_HIP)
-static int ggml_cuda_parse_id(char devName[]) {
+static int ggml_cuda_parse_id(const char * devName) {
    // A list of possible Target IDs can be found under the rocclr/clr repo in device.cpp
    // these values are not stable so this is susceptible to breakage
    // https://github.com/ROCm/clr/blob/amd-staging/rocclr/device/device.cpp
-    int archMajor = 0x0;
-    int archMinor = 0x0;
-    int archNum = GGML_CUDA_CC_OFFSET_AMD;
-    int archLen = strlen(devName);
-    char archName[archLen + 1];
+    // Extract the base architecture model ("gfx908:sramecc+:xnack-" -> "908")
+    std::string archName = devName;
+    archName.erase(0, 3);
+    archName = archName.substr(0, archName.find(':'));

-    // strip leading 'gfx' while copying into our buffer
-    if (archLen > 3) {
-        strcpy(archName, &devName[3]);
-        archLen -= 3;
-    }
+    const auto hex = [](const std::string & s) { return (int) strtoul(s.c_str(), nullptr, 16); };

-    // trim trailing :xnack- or :sramecc- statuses
-    archLen = strcspn(archName, ":");
-    archName[archLen] = '\0';
+    int cc = 0;

    // tease out the version information
-    if (archLen > 8) {
+    if (archName.size() > 8) {
        // versions labeled generic use '-' as delimiter
-        // strip the trailing "-generic" then iterate through what remains
-        if ((strstr(archName, "-generic"))) {
-            archName[archLen - 8] = '\0';
-            char * pch;
-            if ((pch = strtok(archName, "-"))) {
-                archMajor = (int)strtoul(pch, 0, 16);
-                if ((pch = strtok(NULL, "-"))) {
-                    archMinor = 0x10 * (int)strtoul(pch, 0, 16);
-                }
+        if (archName.compare(archName.size() - 8, 8, "-generic") == 0) {
+            archName.resize(archName.size() - 8);
+            const size_t sep = archName.find('-');
+            cc = 0x100 * hex(archName.substr(0, sep));
+            if (sep != std::string::npos) {
+                cc += 0x10 * hex(archName.substr(sep + 1));
            }
        }
-    } else if (archLen >= 3) {
-        // last two digits should be the minor * 0x10 + stepping
-        archMinor = (int)strtoul(&archName[archLen - 2], 0, 16);
-        archName[archLen - 2] = '\0';
-
-        // only the major version remains
-        archMajor = (int)strtoul(archName, 0, 16);
+    } else if (archName.size() >= 3) {
+        // ordinary IDs are the version in hex
+        cc = hex(archName);
    }
-    archNum += archMajor * 0x100;
-    archNum += archMinor;

-    return archNum;
+    return GGML_CUDA_CC_OFFSET_AMD + cc;
 }
 #endif // defined(GGML_USE_HIP)

@@ -457,6 +442,7 @@ const ggml_cuda_device_info & ggml_cuda_info() {
 struct ggml_cuda_pool_leg : public ggml_cuda_pool {
    static const int MAX_BUFFERS = 256;

+    ggml_backend_cuda_context & ctx;
    int device;
    struct ggml_cuda_buffer {
        void * ptr = nullptr;
@@ -466,7 +452,8 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
    ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
    size_t pool_size = 0;

-    explicit ggml_cuda_pool_leg(int device) :
+    ggml_cuda_pool_leg(ggml_backend_cuda_context & ctx, int device) :
+        ctx(ctx),
        device(device) {
    }

@@ -488,6 +475,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
        }
    }

+    // captured CUDA graphs can still use freed buffers, so they must be captured again
+    void invalidate_cuda_graphs() {
+#ifdef USE_CUDA_GRAPH
+        int n_captured = 0;
+        for (const auto & [key, graph] : ctx.cuda_graphs) {
+            n_captured += graph->instances.size();
+        }
+        if (n_captured > 0) {
+            GGML_LOG_DEBUG(GGML_CUDA_NAME " pool[%d]: memory freed, %d captured CUDA graphs must be captured again\n", device, n_captured);
+        }
+        ctx.cuda_graphs.clear();
+#endif // USE_CUDA_GRAPH
+    }
+
    void * alloc(size_t size, size_t * actual_size) override {
 #ifdef DEBUG_CUDA_MALLOC
        int nnz = 0;
@@ -538,6 +539,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
                           device, look_ahead_size/1024.0/1024.0, cached_bytes/1024.0/1024.0);
            CUDA_CHECK(cudaDeviceSynchronize());
            clear_pool();
+            invalidate_cuda_graphs();
            err = ggml_cuda_device_malloc(&ptr, look_ahead_size, device);
            if (err == cudaSuccess) {
                GGML_LOG_DEBUG(GGML_CUDA_NAME " pool[%d]: retry succeeded\n", device);
@@ -566,6 +568,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
        ggml_cuda_set_device(device);
        CUDA_CHECK(cudaFree(ptr));
        pool_size -= size;
+        invalidate_cuda_graphs();
    }
 };

@@ -720,14 +723,15 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
 };
 #endif // defined(GGML_USE_VMM)

-std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int                  device,
-                                                                               [[maybe_unused]] int stream_no) {
+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(ggml_backend_cuda_context & ctx,
+                                                                               int                         device,
+                                                                               [[maybe_unused]] int        stream_no) {
 #if defined(GGML_USE_VMM)
    if (ggml_cuda_info().devices[device].vmm) {
        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
    }
 #endif // defined(GGML_USE_VMM)
-    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
+    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(ctx, device));
 }

 // destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
@@ -1953,6 +1957,61 @@ static bool ggml_cuda_match_shared_expert(const ggml_cgraph * graph, int routed_
    return true;
 }

+// Buffer- and batch-size independent, so graph_optimize produces the same graph topology for every ubatch size (see #29986).
+static bool ggml_cuda_mul_mat_q_fusion_matches(const ggml_tensor * up, const ggml_tensor * gate, const ggml_tensor * glu, const int device) {
+    const ggml_tensor * x_up = up->src[0];
+    const ggml_tensor * y    = up->src[1];
+
+    if (up->op != GGML_OP_MUL_MAT || !ggml_cuda_mmq_fusion_supported(x_up->type) || y->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    // The fused kernel is only compiled for weights whose rows fill whole tiles (no fallback).
+    if (ggml_cuda_mmq_needs_fallback(x_up->ne[1])) {
+        return false;
+    }
+
+    if (!ggml_cuda_should_fuse_mul_mat(up, gate, glu)) {
+        return false;
+    }
+
+    // The fused write-back applies SWIGLU_OAI with the default alpha and limit of ggml_cuda_op_swiglu_oai_single.
+    if (ggml_get_glu_op(glu) == GGML_GLU_OP_SWIGLU_OAI && (ggml_get_op_params_f32(glu, 2) != 1.702f || ggml_get_op_params_f32(glu, 3) != 7.0f)) {
+        return false;
+    }
+
+    // NVIDIA only: the fused write-back relies on the MMA accumulator layout.
+    return turing_mma_available(ggml_cuda_info().devices[device].cc);
+}
+
+static bool ggml_cuda_should_fuse_mul_mat_q(const ggml_tensor * up, const ggml_tensor * gate, const ggml_tensor * glu) {
+    // A hinted matmul can be fused, but its unfused path may be faster (e.g. FWHT for GGML_HINT_SRC0_IS_HADAMARD).
+    if (ggml_get_op_params_i32(up, 1) != GGML_HINT_NONE || ggml_get_op_params_i32(gate, 1) != GGML_HINT_NONE) {
+        return false;
+    }
+
+    // The padding of a compute-buffer weight that is a view cannot be cleared, see ggml_cuda_mul_mat.
+    const auto bad_padding_clear = [](const ggml_tensor * x) {
+        return ggml_backend_buffer_get_usage(x->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE &&
+            ggml_nbytes(x) != ggml_backend_buffer_get_alloc_size(x->buffer, x) && x->view_src;
+    };
+    const int device = ggml_cuda_get_device();
+    if (!ggml_cuda_mul_mat_q_fusion_matches(up, gate, glu, device) || bad_padding_clear(up->src[0]) || bad_padding_clear(gate->src[0])) {
+        return false;
+    }
+
+    // The batch size checks are done here, not in ggml_cuda_mul_mat_q_fusion_matches.
+    const ggml_tensor * x_up = up->src[0];
+    const ggml_tensor * y    = up->src[1];
+    const int           cc   = ggml_cuda_info().devices[device].cc;
+    if (ggml_cuda_should_use_mmvq(x_up->type, cc, y->ne[1]) || !ggml_cuda_should_use_mmq(x_up->type, cc, y->ne[1], 0)) {
+        return false;
+    }
+
+    // The fused kernel needs one block per tile (no stream-k split), so only fuse if the tiling is efficient.
+    return ggml_cuda_mmq_fusion_is_efficient(x_up->type, x_up->ne[1], y->ne[1], y->ne[2]*y->ne[3], device);
+}
+
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_TENSOR_BINARY_OP_LOCALS

@@ -2014,6 +2073,8 @@ static bool ggml_cuda_mul_mat_id_needs_sync(const ggml_tensor * dst, const int c
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];

+    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
+
    if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
        return true;
    }
@@ -2032,7 +2093,7 @@ static bool ggml_cuda_mul_mat_id_needs_sync(const ggml_tensor * dst, const int c
        return false;
    }

-    if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
+    if (ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
        return false;
    }

@@ -2050,6 +2111,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
    GGML_TENSOR_BINARY_OP_LOCALS

    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;

    // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
@@ -2074,7 +2136,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
            return;
        }

-        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
+        if (ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
            ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
            return;
        }
@@ -2548,6 +2610,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
        case GGML_OP_LIGHTNING_INDEXER:
            ggml_cuda_lightning_indexer(ctx, dst);
            break;
+        case GGML_OP_TOP_K_MASK:
+            ggml_cuda_op_top_k_mask(ctx, dst);
+            break;
        default:
            return false;
    }
@@ -2777,16 +2842,14 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
    return res;
 }

-static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
-    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
-
+static void ggml_cuda_graph_update_executable(cudaGraphExec_t & instance, cudaGraph_t cuda_graph) {
 #if CUDART_VERSION >= 12000
    cudaGraphExecUpdateResultInfo result_info;
-    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(instance, cuda_graph, &result_info);
 #else
    cudaGraphNode_t errorNode;
    cudaGraphExecUpdateResult result_info;
-    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &errorNode, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(instance, cuda_graph, &errorNode, &result_info);
 #endif // CUDART_VERSION >= 12000

    if (stat == cudaErrorGraphExecUpdateFailure) {
@@ -2797,13 +2860,39 @@ static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_c
        // The pre-existing graph exec cannot be updated due to violated constraints
        // so instead clear error and re-instantiate
        (void)cudaGetLastError();
-        CUDA_CHECK(cudaGraphExecDestroy(graph->instance));
-        graph->instance = nullptr;
-        CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
+        CUDA_CHECK(cudaGraphExecDestroy(instance));
+        instance = nullptr;
+        CUDA_CHECK(cudaGraphInstantiate(&instance, cuda_graph, nullptr, nullptr, 0));
    } else {
        GGML_ASSERT(stat == cudaSuccess);
    }
 }
+
+// number of ggml nodes per CUDA graph chunk, 0 captures the whole ggml graph as one CUDA graph
+static int ggml_cuda_graph_chunk_size() {
+    static const int chunk_size = [] {
+        const char * env = getenv("GGML_CUDA_GRAPH_CHUNK_SIZE");
+        return env != nullptr ? atoi(env) : 512;
+    }();
+    return chunk_size;
+}
+
+// end the capture of chunk `idx`, update or create its executable and launch it
+static void ggml_cuda_graph_launch_chunk(ggml_backend_cuda_context * cuda_ctx, ggml_cuda_graph * graph, size_t idx) {
+    cudaGraph_t cuda_graph = nullptr;
+    CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_graph));
+
+    if (idx < graph->instances.size()) {
+        ggml_cuda_graph_update_executable(graph->instances[idx], cuda_graph);
+    } else {
+        GGML_ASSERT(idx == graph->instances.size());
+        graph->instances.emplace_back(nullptr);
+        CUDA_CHECK(cudaGraphInstantiate(&graph->instances[idx], cuda_graph, nullptr, nullptr, 0));
+    }
+    CUDA_CHECK(cudaGraphDestroy(cuda_graph));
+
+    CUDA_CHECK(cudaGraphLaunch(graph->instances[idx], cuda_ctx->stream()));
+}
 #endif // USE_CUDA_GRAPH

 static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
@@ -2831,9 +2920,10 @@ static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
        return false;
    }

-    // Only norm/neox shaders have the fusion code
+    // Only norm/neox/multi shaders have the fusion code
    const int mode = ((const int32_t *) rope->op_params)[2];
-    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX) {
+    if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX &&
+        mode != GGML_ROPE_TYPE_MROPE && mode != GGML_ROPE_TYPE_IMROPE) {
        return false;
    }

@@ -4327,6 +4417,18 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                fused_node_count  = 3;
                break;
            }
+
+            if (ggml_cuda_should_fuse_mul_mat_q(up, gate, glu)) {
+                ggml_cuda_mm_fusion_args_host fusion_data{};
+                fusion_data.gate      = gate->src[0];
+                fusion_data.glu_op    = ggml_get_glu_op(glu);
+                fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
+
+                ggml_cuda_mul_mat_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                fused_mul_mat_vec = true;
+                fused_node_count  = 3;
+                break;
+            }
        }
    }

@@ -4557,6 +4659,13 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
    ggml_cuda_concurrent_event * concurrent_event           = nullptr;
    bool                         should_launch_concurrent_events = false;

+#ifdef USE_CUDA_GRAPH
+    ggml_cuda_graph * graph      = use_cuda_graph ? cuda_ctx->cuda_graph(graph_key) : nullptr;
+    const int         chunk_size = ggml_cuda_graph_chunk_size();
+    int               chunk_end  = chunk_size;
+    size_t            n_chunks   = 0;
+#endif // USE_CUDA_GRAPH
+
    const auto try_launch_concurrent_event = [&](const ggml_tensor * node) {
        if (stream_ctx.concurrent_events.find(node) != stream_ctx.concurrent_events.end()) {
            concurrent_event = &stream_ctx.concurrent_events[node];
@@ -4646,6 +4755,15 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud

            for (int i = 0; i < cgraph->n_nodes; i++) {
                ggml_tensor * node = cgraph->nodes[i];
+#ifdef USE_CUDA_GRAPH
+                // fused nodes are skipped together, so a cut never splits a fusion
+                // a concurrent region must join before its capture can end
+                if (use_cuda_graph && chunk_size > 0 && i >= chunk_end && !is_concurrent_event_active) {
+                    ggml_cuda_graph_launch_chunk(cuda_ctx, graph, n_chunks++);
+                    CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
+                    chunk_end = i + chunk_size;
+                }
+#endif // USE_CUDA_GRAPH
                if (is_concurrent_event_active) {
                    GGML_ASSERT(concurrent_event);

@@ -4728,14 +4846,14 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
        }

 #ifdef USE_CUDA_GRAPH
-        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
        if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
-            if (graph->graph != nullptr) {
-                CUDA_CHECK(cudaGraphDestroy(graph->graph));
-                graph->graph = nullptr;
-            }
+            ggml_cuda_graph_launch_chunk(cuda_ctx, graph, n_chunks++);

-            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &graph->graph));
+            // drop the chunks left over from a longer graph
+            for (size_t i = n_chunks; i < graph->instances.size(); ++i) {
+                CUDA_CHECK(cudaGraphExecDestroy(graph->instances[i]));
+            }
+            graph->instances.resize(n_chunks);
            graph_evaluated_or_captured = true; // CUDA graph has been captured

            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
@@ -4747,16 +4865,11 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
        }
    }

-    if (use_cuda_graph) {
-        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
-        if (graph->instance == nullptr) { // Create executable graph from captured graph.
-            CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
+    if (use_cuda_graph && !cuda_graph_update_required) {
+        // the chunks were launched during capture, otherwise launch them here
+        for (cudaGraphExec_t instance : graph->instances) {
+            CUDA_CHECK(cudaGraphLaunch(instance, cuda_ctx->stream()));
        }
-        if (cuda_graph_update_required) { // Update graph executable
-            ggml_cuda_graph_update_executable(cuda_ctx, graph_key);
-        }
-        // Launch graph
-        CUDA_CHECK(cudaGraphLaunch(graph->instance, cuda_ctx->stream()));
 #else
        GGML_UNUSED(graph_key);
        graph_evaluated_or_captured = true;
@@ -4768,7 +4881,7 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
 static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);

-    if (graph->graph == nullptr) {
+    if (graph->instances.empty()) {
        if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_VOLTA) {
            if (!graph->disable_due_to_gpu_arch && ggml_cuda_graph_debug_enabled()) {
                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
@@ -4822,7 +4935,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
                    }
                } else {
                    use_cuda_graph = true;
-                    cuda_graph_update_required = graph->instance == nullptr;
+                    cuda_graph_update_required = graph->instances.empty();
                }
            }
        }
@@ -4912,6 +5025,19 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
            }
        }
        for (int i = 0; i < cgraph->n_nodes; ++i) {
+            // Keep the inputs of the fused gate/up/GLU MMQ kernel alive until the GLU so that they cannot alias its output (under op offload the weights are buffer copies).
+            if (i + 2 < cgraph->n_nodes && cgraph->nodes[i + 2]->op == GGML_OP_GLU) {
+                ggml_tensor * glu  = cgraph->nodes[i + 2];
+                ggml_tensor * gate = glu->src[0];
+                ggml_tensor * up   = glu->src[1];
+                const bool ok = (gate == cgraph->nodes[i] && up == cgraph->nodes[i + 1]) || (gate == cgraph->nodes[i + 1] && up == cgraph->nodes[i]);
+                if (ok && ggml_cuda_mul_mat_q_fusion_matches(up, gate, glu, cuda_ctx->device)) {
+                    add_alloc_deps(i, i + 2);
+                    i += 2;
+                    continue;
+                }
+            }
+
            ggml_cuda_moe_weighted_reduction_match match;
            if (ggml_cuda_match_moe_weighted_reduction(cgraph, i, match)) {
                params->add_alloc_dep(params->user_data, const_cast<ggml_tensor *>(match.experts), match.dst);
@@ -5981,6 +6107,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            return true;
        case GGML_OP_LIGHTNING_INDEXER:
            return ggml_cuda_lightning_indexer_supported(dev_ctx->device, op);
+        case GGML_OP_TOP_K_MASK:
+            return op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_F32;

        default:
            return false;
@@ -1409,14 +1409,14 @@ namespace ggml_cuda_mma {
            tile<16, 16, float, dl_d> & D, const tile<16, 8, nv_bfloat162, dl_ab> & A, const tile<16, 8, nv_bfloat162, dl_ab> & B) {
 #if defined(AMD_WMMA_AVAILABLE)
 #if defined(RDNA4)
-        using bf16x8_t = __attribute__((ext_vector_type(8))) __bf16;
+        using bf16x8_t = __attribute__((ext_vector_type(8))) short;
        using floatx8_t = __attribute__((ext_vector_type(8))) float;
        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
        const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
        const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
 #elif defined(RDNA3)
-        using bf16x16_t = __attribute__((ext_vector_type(16))) __bf16;
+        using bf16x16_t = __attribute__((ext_vector_type(16))) short;
        using floatx8_t = __attribute__((ext_vector_type(8))) float;
        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
        const bf16x16_t& a_frag = reinterpret_cast<const bf16x16_t&>(A.x[0]);
@@ -5,7 +5,9 @@
 #include "mmq.cuh"

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q1_0(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -60,10 +62,10 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            const uint32_t q2_indices = (q2_bits & 0x01) | ((q2_bits & 0x02) << 7) | ((q2_bits & 0x04) << 14) | ((q2_bits & 0x08) << 21);
            const uint32_t q3_indices = (q3_bits & 0x01) | ((q3_bits & 0x02) << 7) | ((q3_bits & 0x04) << 14) | ((q3_bits & 0x08) << 21);

-            const int v0 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q0_indices);
-            const int v1 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q1_indices);
-            const int v2 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q2_indices);
-            const int v3 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q3_indices);
+            const int v0 = ggml_hip_perm(0x000001FF, 0x000001FF, q0_indices);
+            const int v1 = ggml_hip_perm(0x000001FF, 0x000001FF, q1_indices);
+            const int v2 = ggml_hip_perm(0x000001FF, 0x000001FF, q2_indices);
+            const int v3 = ggml_hip_perm(0x000001FF, 0x000001FF, q3_indices);
 #else
            // unpack crumbs into nibble indices
            const int n0 = __byte_perm(0x11100100, 0x11100100, q >> 0); // [0, 1, 4, 5] [ 8,  9, 12, 13]
@@ -116,7 +118,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q2_0(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -162,8 +166,8 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            const uint32_t qx_indices = (q & 0x03) | ((q & 0x0C) << 6) | ((q & 0x30) << 12) | ((q & 0xC0) << 18);
            const uint32_t qy_bits    = q >> 8;
            const uint32_t qy_indices = (qy_bits & 0x03) | ((qy_bits & 0x0C) << 6) | ((qy_bits & 0x30) << 12) | ((qy_bits & 0xC0) << 18);
-            const int qx = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qx_indices);
-            const int qy = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qy_indices);
+            const int      qx         = ggml_hip_perm(0x020100FF, 0x020100FF, qx_indices);
+            const int      qy         = ggml_hip_perm(0x020100FF, 0x020100FF, qy_indices);
 #else
            // unpack even and odd crumbs into byte values
            const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
@@ -205,7 +209,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_0(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -268,7 +274,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_1(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -331,7 +339,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_0(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -411,7 +421,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_1(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -488,8 +500,8 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    }
 }

-template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q8_0(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+template <ggml_type type, int J, bool fallback, bool has_fusion> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q8_0(
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -519,7 +531,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx;
+        const block_q8_0 * bxi = ggml_cuda_mmq_x_row<block_q8_0, has_fusion>(x, x_gate, kbx0, i, stride) + kbx;

 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
        x_qs[i*sram_stride + 0             + txi] = get_int_b2(bxi[0].qs,                   kqsx);
@@ -542,7 +554,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd;
+        const block_q8_0 * bxi = ggml_cuda_mmq_x_row<block_q8_0, has_fusion>(x, x_gate, kbx0, i, stride) + kbxd;

 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
        x_df[i*sram_stride                           + kbxd] = bxi->d;
@@ -555,7 +567,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 // ---------------------------------------------------------------------------------------------

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q2_K(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -616,7 +630,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q3_K(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -728,8 +744,8 @@ static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, co
           ((scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030);  // upper 2 bits
 }

-template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_K(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+template <ggml_type type, int J, bool fallback, bool has_fusion> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_K(
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -757,7 +773,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+        const block_q4_K * bxi = ggml_cuda_mmq_x_row<block_q4_K, has_fusion>(x, x_gate, kbx0, i, stride);
        const int qs0 = get_int_b4(bxi->qs, txi);

 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
@@ -786,7 +802,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
                i = min(i, i_max);
            }

-            const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+            const block_q4_K * bxi = ggml_cuda_mmq_x_row<block_q4_K, has_fusion>(x, x_gate, kbx0, i, stride);

            const int * scales = (const int *) bxi->scales;
            const int ksc = threadIdx.x % 2;
@@ -814,7 +830,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+        const block_q4_K * bxi = ggml_cuda_mmq_x_row<block_q4_K, has_fusion>(x, x_gate, kbx0, i, stride);

        x_dm[i] = bxi->dm;
    }
@@ -827,7 +843,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / (QI4_K/8);
+        const block_q4_K * bxi = ggml_cuda_mmq_x_row<block_q4_K, has_fusion>(x, x_gate, kbx0, i, stride) + (threadIdx.x % (MMQ_TILE_NE_K/8)) / (QI4_K/8);

        const int * scales = (const int *) bxi->scales;

@@ -839,8 +855,8 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }

-template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_K(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+template <ggml_type type, int J, bool fallback, bool has_fusion> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q5_K(
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -868,7 +884,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+        const block_q5_K * bxi = ggml_cuda_mmq_x_row<block_q5_K, has_fusion>(x, x_gate, kbx0, i, stride);
        const int ky = QR5_K*txi;

        const int ql = get_int_b4(bxi->qs, txi);
@@ -909,7 +925,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
                i = min(i, i_max);
            }

-            const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+            const block_q5_K * bxi = ggml_cuda_mmq_x_row<block_q5_K, has_fusion>(x, x_gate, kbx0, i, stride);

            const int * scales = (const int *) bxi->scales;
            const int ksc = threadIdx.x % 2;
@@ -937,7 +953,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+        const block_q5_K * bxi = ggml_cuda_mmq_x_row<block_q5_K, has_fusion>(x, x_gate, kbx0, i, stride);

        x_dm[i] = bxi->dm;
    }
@@ -951,7 +967,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+        const block_q5_K * bxi = ggml_cuda_mmq_x_row<block_q5_K, has_fusion>(x, x_gate, kbx0, i, stride);

        const int * scales = (const int *) bxi->scales;

@@ -963,8 +979,8 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }

-template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q6_K(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+template <ggml_type type, int J, bool fallback, bool has_fusion> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q6_K(
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -993,7 +1009,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
+        const block_q6_K * bxi = ggml_cuda_mmq_x_row<block_q6_K, has_fusion>(x, x_gate, kbx0, i, stride);

        const int ql = get_int_b2(bxi->ql, txi);
        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
@@ -1023,7 +1039,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
+        const block_q6_K * bxi = ggml_cuda_mmq_x_row<block_q6_K, has_fusion>(x, x_gate, kbx0, i, stride);

 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
        x_df[i*sram_stride]                     = bxi->d;
@@ -1041,7 +1057,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4;
+        const block_q6_K * bxi = ggml_cuda_mmq_x_row<block_q6_K, has_fusion>(x, x_gate, kbx0, i, stride) + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4;

 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
        x_sc[i*sram_stride + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8));
@@ -1054,7 +1070,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 // ---------------------------------------------------------------------------------------------

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq1_s(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1116,7 +1134,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xxs(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1180,7 +1200,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_xs(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1245,7 +1267,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq2_s(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1313,7 +1337,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_xxs(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1377,7 +1403,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq3_s(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1445,8 +1473,8 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    }
 }

-template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_xs(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+template <ggml_type type, int J, bool fallback, bool has_fusion> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_xs(
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1473,7 +1501,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride;
+        const block_iq4_xs * bxi = ggml_cuda_mmq_x_row<block_iq4_xs, has_fusion>(x, x_gate, kbx0, i, stride);

        const int aux_q4 = get_int_b4(bxi->qs, kqsx);
        const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl);
@@ -1497,7 +1525,7 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
            i = min(i, i_max);
        }

-        const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride;
+        const block_iq4_xs * bxi = ggml_cuda_mmq_x_row<block_iq4_xs, has_fusion>(x, x_gate, kbx0, i, stride);

        const float d = __half2float(bxi->d);

@@ -1513,7 +1541,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_iq4_nl(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1582,7 +1612,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 // ---------------------------------------------------------------------------------------------

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q8) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q8);
@@ -1649,7 +1681,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_mxfp4_fp4(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q4) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q4);
@@ -1691,7 +1725,9 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
 }

 template <ggml_type type, int J, bool fallback, ggml_prec prec_src1> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kb0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kb0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1) / warp_size;
    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback, prec_src1);
@@ -1747,7 +1783,9 @@ template <ggml_type type, int J, bool fallback, ggml_prec prec_src1> static __de
 }

 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_nvfp4_nvfp4(
-        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+        const char * __restrict__ x, const char * __restrict__ x_gate, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    GGML_UNUSED(x_gate);
+
    constexpr int warp_size       = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps          = ggml_cuda_mmq_get_nthreads(type, J, fallback, GGML_PREC_Q4) / warp_size;
    constexpr int I               = ggml_cuda_mmq_get_I(type, J, fallback, GGML_PREC_Q4);
@@ -131,12 +131,51 @@ static ggml_prec ggml_cuda_mmq_get_prec_src1(const ggml_tensor * src0, const ggm
    return GGML_PREC_Q4;
 }

+// Tile width in ne11 (dense) / ne12 (MoE) direction with the fewest tiles that fits into shared memory, 0 if there is none.
+static int ggml_cuda_mmq_get_J_best(
+        const ggml_type type, const int64_t ncols, const bool fallback, const int cc, const size_t smpbo, const ggml_prec prec_src1,
+        int * nthreads_best = nullptr) {
+    int J_best        = 0;
+    int ntiles_J_best = INT_MAX;
+
+    for (int J = 8; J <= 128 && ntiles_J_best > 1; J += 8) {
+        const ggml_cuda_mmq_config config = ggml_cuda_mmq_get_config(type, J, fallback, cc, prec_src1);
+        if (config.type == GGML_TYPE_COUNT) {
+            continue;
+        }
+
+        if (mmq_get_nbytes_shared(config, cc) > smpbo) {
+            continue;
+        }
+
+        const int ntiles_x = (ncols + config.J - 1) / config.J;
+
+        if (ntiles_x < ntiles_J_best) {
+            J_best = J;
+            if (nthreads_best) {
+                *nthreads_best = config.nthreads;
+            }
+            ntiles_J_best = ntiles_x;
+        }
+    }
+    return J_best;
+}
+
 void ggml_cuda_mul_mat_q(
-        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+        const ggml_cuda_mm_fusion_args_host * fusion) {
    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.

+    // Fused gate/up/GLU: src0 is the up weight, dst the GLU output.
+    const ggml_tensor * gate = fusion ? fusion->gate : nullptr;
+    if (fusion) {
+        // TODO MoE support
+        GGML_ASSERT(gate && !ids && !fusion->x_bias && !fusion->gate_bias && !fusion->x_scale && !fusion->gate_scale);
+        GGML_ASSERT(gate->type == src0->type && ggml_are_same_shape(gate, src0) && ggml_are_same_stride(gate, src0));
+    }
+
    GGML_TENSOR_BINARY_OP_LOCALS;

    cudaStream_t stream = ctx.stream();
@@ -158,14 +197,16 @@ void ggml_cuda_mul_mat_q(
    const float * src1_d = (const float *) src1->data;
    float       *  dst_d = (float       *)  dst->data;

-    // If src0 is a temporary compute buffer, clear any potential padding.
-    if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
-        const size_t size_data  = ggml_nbytes(src0);
-        const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
-        if (size_alloc > size_data) {
-            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
-            GGML_ASSERT(!src0->view_src);
-            CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
+    // If src0 (or the fused gate weight) is a temporary compute buffer, clear any potential padding.
+    for (const ggml_tensor * t : { src0, gate }) {
+        if (t && ggml_backend_buffer_get_usage(t->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
+            const size_t size_data  = ggml_nbytes(t);
+            const size_t size_alloc = ggml_backend_buffer_get_alloc_size(t->buffer, t);
+            if (size_alloc > size_data) {
+                GGML_ASSERT(ggml_is_contiguously_allocated(t));
+                GGML_ASSERT(!t->view_src);
+                CUDA_CHECK(cudaMemsetAsync((char *) t->data + size_data, 0, size_alloc - size_data, stream));
+            }
        }
    }

@@ -180,7 +221,9 @@ void ggml_cuda_mul_mat_q(

    const bool fallback = ggml_cuda_mmq_needs_fallback(ne01);

-    const ggml_prec prec_src1 = ggml_cuda_mmq_get_prec_src1(src0, dst, cc);
+    // With fusion dst is the GLU output, not a MUL_MAT; the fused types always quantize src1 to Q8_1.
+    // TODO support for Q4 precision if both tensors allow it
+    const ggml_prec prec_src1 = gate ? GGML_PREC_Q8 : ggml_cuda_mmq_get_prec_src1(src0, dst, cc);

    const bool use_native_fp4 = prec_src1 == GGML_PREC_Q4;
    const size_t y_block_size       = use_native_fp4 ? sizeof(block_fp4_mmq) : sizeof(block_q8_1_mmq);
@@ -201,26 +244,7 @@ void ggml_cuda_mul_mat_q(
            }
        }

-        int ntiles_J_best = INT_MAX;
-
-        for (int J = 8; J <= 128 && ntiles_J_best > 1; J += 8) {
-            const ggml_cuda_mmq_config config = ggml_cuda_mmq_get_config(src0->type, J, fallback, cc, prec_src1);
-            if (config.type == GGML_TYPE_COUNT) {
-                continue;
-            }
-
-            if (mmq_get_nbytes_shared(config, cc) > smpbo) {
-                continue;
-            }
-
-            const int ntiles_x = (ncols_opt + config.J - 1) / config.J;
-
-            if (ntiles_x < ntiles_J_best) {
-                J_best = J;
-                nthreads_best = config.nthreads;
-                ntiles_J_best = ntiles_x;
-            }
-        }
+        J_best = ggml_cuda_mmq_get_J_best(src0->type, ncols_opt, fallback, cc, smpbo, prec_src1, &nthreads_best);
    }
    GGML_ASSERT(J_best > 0);

@@ -262,13 +286,20 @@ void ggml_cuda_mul_mat_q(
                                ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
        const int64_t s13 = ne12*s12;

+        ggml_cuda_mm_fusion_args_device fusion_device = {};
+        if (gate) {
+            fusion_device.gate      = gate->data;
+            fusion_device.glu_op    = fusion->glu_op;
+            fusion_device.glu_limit = fusion->glu_limit;
+        }
+
        const mmq_args args = {
            src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d,
            src0->type == GGML_TYPE_NVFP4 && use_native_fp4 ? src1_scale.ptr : nullptr,
            ne00, ne01, ne1, s01, ne11, s1,
            ne02, ne12, s02, s12, s2,
            ne03, ne13, s03, s13, s3,
-            ne1, J_best};
+            ne1, J_best, fusion_device};
        ggml_cuda_mul_mat_q_switch_type(ctx, args, stream, prec_src1);
        return;
    }
@@ -485,3 +516,22 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t

    return (!GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }
+
+bool ggml_cuda_mmq_fusion_is_efficient(const ggml_type type, const int64_t nrows_x, const int64_t ncols, const int64_t nchannels, const int device) {
+    const int    cc    = ggml_cuda_info().devices[device].cc;
+    const int    nsm   = ggml_cuda_info().devices[device].nsm;
+    const size_t smpbo = ggml_cuda_info().devices[device].smpbo;
+
+    // Same tile width as ggml_cuda_mul_mat_q; the fused kernel has no fallback, quantizes src1 to Q8_1 and writes I/2 output rows per tile.
+    const bool fallback = false;
+    const int  J_best   = ggml_cuda_mmq_get_J_best(type, ncols, fallback, cc, smpbo, GGML_PREC_Q8);
+    if (J_best == 0) {
+        return false;
+    }
+    const int64_t nrows_tile = ggml_cuda_mmq_get_config(type, J_best, fallback, cc, GGML_PREC_Q8).I / 2;
+    const int64_t ntiles     = (ncols + J_best - 1)/J_best * ((nrows_x + nrows_tile - 1)/nrows_tile) * nchannels;
+    // Same condition as launch_mul_mat_q uses to run the stream-k kernel with one block per tile.
+    const int64_t tiles_nwaves             = (ntiles + nsm - 1) / nsm;
+    const int64_t tiles_efficiency_percent = 100 * ntiles / (nsm*tiles_nwaves);
+    return tiles_efficiency_percent >= 90;
+}
@@ -1,6 +1,7 @@
 #pragma once

 #include "common.cuh"
+#include "unary.cuh"

 #include <climits>
 #include <cstdint>
@@ -10,10 +11,12 @@
 #define MMQ_ITER_K_FP4         512
 #define MMQ_NWARPS               8

-typedef void (*ggml_cuda_mmq_load_tiles_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride);
+typedef void (*ggml_cuda_mmq_load_tiles_t)(const char * __restrict__ x, const char * __restrict__ x_gate, int * x_tile,
+    const int kbx0, const int i_max, const int stride);
 typedef void (*ggml_cuda_mmq_vec_dot_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00);
 typedef void (*ggml_cuda_mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted,
-    float * __restrict__ dst, const float * __restrict__ y_scale, const int stride, const int i_max, const int j_max);
+    float * __restrict__ dst, const float * __restrict__ y_scale, const int stride, const int i_max, const int j_max,
+    const ggml_glu_op glu_op, const float glu_limit);

 enum mmq_q8_1_ds_layout {
    MMQ_Q8_1_DS_LAYOUT_D4,
@@ -426,15 +429,41 @@ static __host__ int ggml_cuda_mmq_get_nbytes_shared_x(const ggml_cuda_mmq_config

 // ------------------------------------------------------------

+// Weight types with a fused gate/up/GLU path (see ggml_cuda_mmq_get_util_funcs).
+static constexpr __host__ __device__ bool ggml_cuda_mmq_fusion_supported(const ggml_type type) {
+    return type == GGML_TYPE_Q4_K || type == GGML_TYPE_Q5_K || type == GGML_TYPE_Q6_K || type == GGML_TYPE_Q8_0 || type == GGML_TYPE_IQ4_XS;
+}
+
+// Fused gate/up/GLU: the weight tile alternates chunks of this many up and gate rows, so rows r and r + 8 of an MMA tile (same thread) pair up.
+static constexpr int GGML_CUDA_MMQ_FUSION_ROWS = 8;
+
+// Source of weight tile row i; with has_fusion the up or gate row that the interleaved tile holds there.
+template <typename block_t, bool has_fusion>
+static __device__ __forceinline__ const block_t * ggml_cuda_mmq_x_row(
+        const char * __restrict__ x, const char * __restrict__ x_gate, const int kbx0, const int i, const int stride) {
+    if constexpr (has_fusion) {
+        constexpr int R = GGML_CUDA_MMQ_FUSION_ROWS;
+        const char * x_src = (i / R) % 2 == 0 ? x : x_gate;
+        return (const block_t *) x_src + kbx0 + ((i / (2*R))*R + i % R)*stride;
+    } else {
+        GGML_UNUSED(x_gate);
+        return (const block_t *) x + kbx0 + i*stride;
+    }
+}
+
 #include "mmq-load-tiles.cuh"
 #include "mmq-vec-dot.cuh"

 template <ggml_type type, int J, bool fallback, ggml_prec prec_src1> static __device__ __forceinline__ void ggml_cuda_mmq_write_back_dp4a(
        const float * __restrict__ sum, const int32_t * __restrict__ ids_dst, float * __restrict__ dst,
-        const float * __restrict__ y_scale, const int stride, const int i_max, const int j_max) {
+        const float * __restrict__ y_scale, const int stride, const int i_max, const int j_max,
+        const ggml_glu_op glu_op, const float glu_limit) {
+    GGML_UNUSED_VARS(glu_op, glu_limit);
+
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
    constexpr int nwarps    = ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1) / warp_size;
    constexpr int I         = ggml_cuda_mmq_get_I(type, J, fallback, prec_src1);
+    static_assert(I % warp_size == 0, "DP4A MMQ tile height must be a multiple of the physical warp size");

    const bool y_scale_used = y_scale != nullptr;

@@ -468,16 +497,19 @@ template <ggml_type type, int J, bool fallback, ggml_prec prec_src1> static __de
    }
 }

-template<ggml_type type, int J, bool fallback, ggml_prec prec_src1>
+// has_fusion: accumulators l and l + ne/2 hold the up and gate values of one output element, the GLU of the two is written.
+template<ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion>
 static __device__ __forceinline__ void ggml_cuda_mmq_write_back_mma(
            const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst,
-            const float * __restrict__ y_scale, const int stride, const int i_max, const int j_max) {
+            const float * __restrict__ y_scale, const int stride, const int i_max, const int j_max,
+            const ggml_glu_op glu_op, const float glu_limit) {

 #if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
 #else
    typedef tile<16,  8, int> tile_C;
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    static_assert(!has_fusion || tile_C::I == 2*GGML_CUDA_MMQ_FUSION_ROWS, "each C tile must hold one chunk of up rows and one of gate rows");

    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback, prec_src1);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -491,20 +523,23 @@ static __device__ __forceinline__ void ggml_cuda_mmq_write_back_mma(
 #pragma unroll
        for (int n = 0; n < ntx; ++n) {
 #pragma unroll
-            for (int l = 0; l < tile_C::ne; ++l) {
+            for (int l = 0; l < (has_fusion ? tile_C::ne/2 : tile_C::ne); ++l) {
                const int j = j0 + (threadIdx.y % ntx) * tile_C::J + tile_C::get_j(l);

                if (j > j_max) {
                    continue;
                }

-                const int i = i0 + n*tile_C::I + tile_C::get_i(l);
+                const int i = (i0 + n*tile_C::I)/(has_fusion ? 2 : 1) + tile_C::get_i(l);

                if (fallback && i > i_max) {
                    continue;
                }

-                if constexpr (type == GGML_TYPE_NVFP4) {
+                if constexpr (has_fusion) {
+                    const int sum_index = (j0/tile_C::J + n)*tile_C::ne + l;
+                    dst[ids_dst[j]*stride + i] = ggml_cuda_op_glu_single(glu_op, sum[sum_index + tile_C::ne/2], sum[sum_index], glu_limit);
+                } else if constexpr (type == GGML_TYPE_NVFP4) {
                    if (y_scale_used) {
                        dst[ids_dst[j]*stride + i] = y_scale[j] * sum[(j0/tile_C::J + n)*tile_C::ne + l];
                    } else {
@@ -517,6 +552,8 @@ static __device__ __forceinline__ void ggml_cuda_mmq_write_back_mma(
            }
        }
    }
+
+    GGML_UNUSED_VARS(y_scale_used, glu_op, glu_limit);
 }

 // -------------------------------------------------------------------------------------------------------------------------------------
@@ -533,8 +570,9 @@ struct ggml_cuda_mmq_util_funcs {
        vdr(vdr), load_tiles(load_tiles), vec_dot(vec_dot), write_back(write_back) {}
 };

-template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
+template <ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion>
 static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_funcs() {
+    static_assert(!has_fusion || ggml_cuda_mmq_fusion_supported(type), "no fused loader for this type");
    if (!ggml_cuda_mmq_get_config(type, J, fallback, prec_src1).use_mma_data_layout()) {
        switch (type) {
            case GGML_TYPE_Q1_0:
@@ -576,7 +614,7 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
            case GGML_TYPE_Q8_0:
                return ggml_cuda_mmq_util_funcs(
                    VDR_Q8_0_Q8_1_MMQ,
-                    ggml_cuda_mmq_load_tiles_q8_0<type, J, fallback>,
+                    ggml_cuda_mmq_load_tiles_q8_0<type, J, fallback, false>,
                    ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a<type, J, fallback>,
                    ggml_cuda_mmq_write_back_dp4a<type, J, fallback, prec_src1>);
 // ---------------------------------------------------------------------------------------------
@@ -595,19 +633,19 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
            case GGML_TYPE_Q4_K:
                return ggml_cuda_mmq_util_funcs(
                    VDR_Q4_K_Q8_1_MMQ,
-                    ggml_cuda_mmq_load_tiles_q4_K<type, J, fallback>,
+                    ggml_cuda_mmq_load_tiles_q4_K<type, J, fallback, false>,
                    ggml_cuda_mmq_vec_dot_q4_K_q8_1_dp4a<type, J, fallback>,
                    ggml_cuda_mmq_write_back_dp4a<type, J, fallback, prec_src1>);
            case GGML_TYPE_Q5_K:
                return ggml_cuda_mmq_util_funcs(
                    VDR_Q5_K_Q8_1_MMQ,
-                    ggml_cuda_mmq_load_tiles_q5_K<type, J, fallback>,
+                    ggml_cuda_mmq_load_tiles_q5_K<type, J, fallback, false>,
                    ggml_cuda_mmq_vec_dot_q5_K_q8_1_dp4a<type, J, fallback>,
                    ggml_cuda_mmq_write_back_dp4a<type, J, fallback, prec_src1>);
            case GGML_TYPE_Q6_K:
                return ggml_cuda_mmq_util_funcs(
                    VDR_Q6_K_Q8_1_MMQ,
-                    ggml_cuda_mmq_load_tiles_q6_K<type, J, fallback>,
+                    ggml_cuda_mmq_load_tiles_q6_K<type, J, fallback, false>,
                    ggml_cuda_mmq_vec_dot_q6_K_q8_1_dp4a<type, J, fallback>,
                    ggml_cuda_mmq_write_back_dp4a<type, J, fallback, prec_src1>);
 // ---------------------------------------------------------------------------------------------
@@ -650,7 +688,7 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
            case GGML_TYPE_IQ4_XS:
                return ggml_cuda_mmq_util_funcs(
                    VDR_IQ4_XS_Q8_1_MMQ,
-                    ggml_cuda_mmq_load_tiles_iq4_xs<type, J, fallback>,
+                    ggml_cuda_mmq_load_tiles_iq4_xs<type, J, fallback, false>,
                    ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a<type, J, fallback>,
                    ggml_cuda_mmq_write_back_dp4a<type, J, fallback, prec_src1>);
            case GGML_TYPE_IQ4_NL:
@@ -687,7 +725,7 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
                    -1,
                    ggml_cuda_mmq_load_tiles_mxfp4_fp4<type, J, fallback>,
                    ggml_cuda_mmq_vec_dot_fp4_fp4_mma<type, J, fallback>,
-                    ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                    ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
            }
            break;
        case GGML_TYPE_NVFP4:
@@ -696,7 +734,7 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
                    -1,
                    ggml_cuda_mmq_load_tiles_nvfp4_nvfp4<type, J, fallback>,
                    ggml_cuda_mmq_vec_dot_fp4_fp4_mma<type, J, fallback>,
-                    ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                    ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
            }
            break;
        default:
@@ -712,136 +750,136 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
                -1,
                ggml_cuda_mmq_load_tiles_q1_0<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q2_0:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_q2_0<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q4_0:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_q4_0<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_DS4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q4_1:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_q4_1<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma<type, J, fallback>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q5_0:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_q5_0<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q5_1:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_q5_1<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma<type, J, fallback>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q8_0:
            return ggml_cuda_mmq_util_funcs(
                -1,
-                ggml_cuda_mmq_load_tiles_q8_0<type, J, fallback>,
+                ggml_cuda_mmq_load_tiles_q8_0<type, J, fallback, has_fusion>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, has_fusion>);
 // ---------------------------------------------------------------------------------------------
        case GGML_TYPE_Q2_K:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_q2_K<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q2_K_q8_1_mma<type, J, fallback>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q3_K:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_q3_K<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma<type, J, fallback, prec_src1>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_Q4_K:
            return ggml_cuda_mmq_util_funcs(
                -1,
-                ggml_cuda_mmq_load_tiles_q4_K<type, J, fallback>,
+                ggml_cuda_mmq_load_tiles_q4_K<type, J, fallback, has_fusion>,
                ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma<type, J, fallback>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, has_fusion>);
        case GGML_TYPE_Q5_K:
            return ggml_cuda_mmq_util_funcs(
                -1,
-                ggml_cuda_mmq_load_tiles_q5_K<type, J, fallback>,
+                ggml_cuda_mmq_load_tiles_q5_K<type, J, fallback, has_fusion>,
                ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma<type, J, fallback>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, has_fusion>);
        case GGML_TYPE_Q6_K:
            return ggml_cuda_mmq_util_funcs(
                -1,
-                ggml_cuda_mmq_load_tiles_q6_K<type, J, fallback>,
+                ggml_cuda_mmq_load_tiles_q6_K<type, J, fallback, has_fusion>,
                ggml_cuda_mmq_vec_dot_q6_K_q8_1_mma<type, J, fallback>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, has_fusion>);
 // ---------------------------------------------------------------------------------------------
        case GGML_TYPE_IQ1_S:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_iq1_s<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_1_q8_1_mma<type, J, fallback>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_IQ2_XXS:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_iq2_xxs<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_IQ2_XS:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_iq2_xs<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma<type, J, fallback, prec_src1>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_IQ2_S:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_iq2_s<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma<type, J, fallback, prec_src1>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_IQ3_XXS:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_iq3_xxs<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_IQ3_S:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_iq3_s<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_IQ4_XS:
            return ggml_cuda_mmq_util_funcs(
                -1,
-                ggml_cuda_mmq_load_tiles_iq4_xs<type, J, fallback>,
+                ggml_cuda_mmq_load_tiles_iq4_xs<type, J, fallback, has_fusion>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, has_fusion>);
        case GGML_TYPE_IQ4_NL:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_iq4_nl<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
 // ---------------------------------------------------------------------------------------------
        case GGML_TYPE_MXFP4:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_mxfp4<type, J, fallback>,
                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        case GGML_TYPE_NVFP4:
            return ggml_cuda_mmq_util_funcs(
                -1,
                ggml_cuda_mmq_load_tiles_nvfp4<type, J, fallback, prec_src1>,
                ggml_cuda_mmq_vec_dot_q8_0_16_q8_1_mma<type, J, fallback, prec_src1>,
-                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1>);
+                ggml_cuda_mmq_write_back_mma<type, J, fallback, prec_src1, false>);
        default:
            return ggml_cuda_mmq_util_funcs(1, nullptr, nullptr, nullptr);
    }
@@ -849,41 +887,44 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func

 template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
 static constexpr __device__ int ggml_cuda_mmq_get_vdr() {
-    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1>().vdr;
+    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1, false>().vdr;
 }

-template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
+template <ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion>
 static constexpr __device__ ggml_cuda_mmq_load_tiles_t ggml_cuda_mmq_get_load_tiles() {
-    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1>().load_tiles;
+    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1, has_fusion>().load_tiles;
 }

 template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
 static constexpr __device__ ggml_cuda_mmq_vec_dot_t ggml_cuda_mmq_get_vec_dot() {
-    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1>().vec_dot;
+    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1, false>().vec_dot;
 }

-template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
+template <ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion>
 static constexpr __device__ ggml_cuda_mmq_write_back_t ggml_cuda_mmq_get_write_back() {
-    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1>().write_back;
+    return ggml_cuda_mmq_get_util_funcs<type, J, fallback, prec_src1, has_fusion>().write_back;
 }

 // ---------------------------------------------------------------------------------------------

-template <ggml_type type, int J, bool fallback, bool fixup, ggml_prec prec_src1>
+// has_fusion: x and x_gate are the up and gate weights.
+template <ggml_type type, int J, bool fallback, bool fixup, ggml_prec prec_src1, bool has_fusion>
 static __device__ __forceinline__ void mul_mat_q_process_tile(
-        const char * __restrict__ x, const int offset_x, const int * __restrict__ y,
+        const char * __restrict__ x, const char * __restrict__ x_gate, const int offset_x, const int * __restrict__ y,
        const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-        const float * __restrict__ y_scale,
+        const float * __restrict__ y_scale, const ggml_glu_op glu_op, const float glu_limit,
        const int stride_row_x, const int ncols_y, const int stride_col_dst,
        const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) {
+    static_assert(!(fixup && has_fusion), "the fused gate/up path does not split K");
+    static_assert(!(fallback && has_fusion), "the fused gate/up path needs nrows_x % 128 == 0");

    constexpr int              warp_size  = ggml_cuda_get_physical_warp_size();
    constexpr int              nwarps     = ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1) / warp_size;
    constexpr int              qk         = ggml_cuda_type_traits<type>::qk;
    constexpr int              I          = ggml_cuda_mmq_get_I(type, J, fallback, prec_src1);
-    constexpr ggml_cuda_mmq_load_tiles_t load_tiles = ggml_cuda_mmq_get_load_tiles<type, J, fallback, prec_src1>();
+    constexpr ggml_cuda_mmq_load_tiles_t load_tiles = ggml_cuda_mmq_get_load_tiles<type, J, fallback, prec_src1, has_fusion>();
    constexpr ggml_cuda_mmq_vec_dot_t    vec_dot    = ggml_cuda_mmq_get_vec_dot<type, J, fallback, prec_src1>();
-    constexpr ggml_cuda_mmq_write_back_t write_back = ggml_cuda_mmq_get_write_back<type, J, fallback, prec_src1>();
+    constexpr ggml_cuda_mmq_write_back_t write_back = ggml_cuda_mmq_get_write_back<type, J, fallback, prec_src1, has_fusion>();

    extern __shared__ int data_mul_mat_q[];
    int * tile_y = data_mul_mat_q + J;
@@ -906,7 +947,7 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
    constexpr int sz = sizeof(block_q8_1_mmq) / sizeof(int);

    for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
-        load_tiles(x, tile_x, offset_x + kb0, tile_x_max_i, stride_row_x);
+        load_tiles(x, x_gate, tile_x, offset_x + kb0, tile_x_max_i, stride_row_x);
        {
            const int * by0 = y + ncols_y * (kb0 * qk / ne_block) * sz;
 #pragma unroll
@@ -941,28 +982,34 @@ static __device__ __forceinline__ void mul_mat_q_process_tile(
    }

    if (fixup) {
-        write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(J*I), y_scale, I, I, J);
+        write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(J*I), y_scale, I, I, J, glu_op, glu_limit);
    } else {
-        write_back(sum, ids_dst, dst, y_scale, stride_col_dst, tile_x_max_i, tile_y_max_j);
+        write_back(sum, ids_dst, dst, y_scale, stride_col_dst, tile_x_max_i, tile_y_max_j, glu_op, glu_limit);
    }
 }


 // The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598

-template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
+template <ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion>
 __launch_bounds__(ggml_cuda_mmq_get_nthreads(type, J, fallback, prec_src1), ggml_cuda_mmq_get_occupancy(type, J, fallback, prec_src1))
 static __global__ void mul_mat_q(
-        const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
+        const char * __restrict__ x, const char * __restrict__ x_gate, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-        const float * __restrict__ y_scale,
+        const float * __restrict__ y_scale, const ggml_glu_op glu_op, const float glu_limit,
        const uint3 blocks_per_ne00, const int nrows_x, const int ncols_dst, const int stride_row_x, const int ncols_y, const int stride_col_dst,
        const uint3 channel_ratio, const uint3 nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
        const uint3 sample_ratio, const uint3 nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
        const uint3 ntx) {

+#ifdef TURING_MMA_AVAILABLE
+    constexpr bool fusion_available = true;
+#else
+    constexpr bool fusion_available = false;
+#endif // TURING_MMA_AVAILABLE
+
    // Skip unused template specializations for faster compilation:
-    if (ggml_cuda_mmq_get_config(type, J, fallback, prec_src1).type == GGML_TYPE_COUNT) {
+    if (ggml_cuda_mmq_get_config(type, J, fallback, prec_src1).type == GGML_TYPE_COUNT || (has_fusion && !fusion_available)) {
        NO_DEVICE_CODE;
        return;
    }
@@ -990,12 +1037,16 @@ static __global__ void mul_mat_q(
    }
    __syncthreads();

-    if constexpr (!ggml_cuda_mmq_get_stream_k(type, J, fallback, prec_src1)) {
+    // The fused kernel needs the whole K range of a tile in one block, so it always uses tiling.
+    if constexpr (has_fusion || !ggml_cuda_mmq_get_stream_k(type, J, fallback, prec_src1)) {
+        // The fused kernel writes the GLU of I/2 up rows and I/2 gate rows per tile.
+        constexpr int I_dst = has_fusion ? I/2 : I;
+
        const uint2 tmp2 = fast_div_modulo(blockIdx.z, nchannels_y);
        const int wt = tmp2.x;
        const int zt = tmp2.y;
-        const int jt = blockIdx.y;
-        const int it = blockIdx.x;
+        const int jt = blockIdx.x;
+        const int it = blockIdx.y;

        // Defaults for regular matrix multiplication:
        int col_low    = 0;
@@ -1040,22 +1091,22 @@ static __global__ void mul_mat_q(
        }

        offset_y   += (col_low + jt*J)*(sizeof(block_q8_1_mmq)/sizeof(int));
-        offset_dst += it*I;
+        offset_dst += it*I_dst;
        const float * y_scale_tile = nullptr;
        if constexpr (type == GGML_TYPE_NVFP4) {
            offset_y_scale += col_low + jt*J;
            y_scale_tile = y_scale ? y_scale + offset_y_scale : nullptr;
        }

-        const int tile_x_max_i = nrows_x  - it*I - 1;
+        const int tile_x_max_i = nrows_x  - it*I_dst - 1;
        const int tile_y_max_j = col_diff - jt*J - 1;

-        const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x;
+        const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I_dst*stride_row_x;

        constexpr bool fixup = false;
-        mul_mat_q_process_tile<type, J, fallback, fixup, prec_src1>
-            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, y_scale_tile,
-             stride_row_x, ncols_y, stride_col_dst,
+        mul_mat_q_process_tile<type, J, fallback, fixup, prec_src1, has_fusion>
+            (x, x_gate, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, y_scale_tile,
+             glu_op, glu_limit, stride_row_x, ncols_y, stride_col_dst,
             tile_x_max_i, tile_y_max_j, 0, blocks_per_ne00.z);
        return;
    }
@@ -1147,9 +1198,9 @@ static __global__ void mul_mat_q(
        const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x;

        constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
-        mul_mat_q_process_tile<type, J, fallback, fixup, prec_src1>
-            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, y_scale_tile,
-             stride_row_x, ncols_y, stride_col_dst,
+        mul_mat_q_process_tile<type, J, fallback, fixup, prec_src1, false>
+            (x, x_gate, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, y_scale_tile,
+             glu_op, glu_limit, stride_row_x, ncols_y, stride_col_dst,
             tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);

        kbc += blocks_per_ne00.z;
@@ -1231,9 +1282,9 @@ static __global__ void mul_mat_q(
    const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*I*stride_row_x;

    constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
-    mul_mat_q_process_tile<type, J, fallback, fixup, prec_src1>
-        (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, y_scale_tile,
-         stride_row_x, ncols_y, stride_col_dst,
+    mul_mat_q_process_tile<type, J, fallback, fixup, prec_src1, false>
+        (x, x_gate, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, y_scale_tile,
+         glu_op, glu_limit, stride_row_x, ncols_y, stride_col_dst,
         tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }

@@ -1383,6 +1434,7 @@ struct mmq_args {
    int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
    int64_t ncols_max;
    int J_best; // Tile width in ne11(dense)/ne12(MoE) direction to use for optimal performance.
+    ggml_cuda_mm_fusion_args_device fusion = {}; // gate != nullptr selects the fused gate/up/GLU path
 };

 static size_t mmq_get_nbytes_shared(const ggml_cuda_mmq_config & config, const int cc) {
@@ -1392,7 +1444,7 @@ static size_t mmq_get_nbytes_shared(const ggml_cuda_mmq_config & config, const i
    return nbs_ids + nbs_x + GGML_PAD(nbs_y, config.nthreads*sizeof(int));
 }

-template <ggml_type type, int J, bool fallback, ggml_prec prec_src1>
+template <ggml_type type, int J, bool fallback, ggml_prec prec_src1, bool has_fusion>
 static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
    const int id = ggml_cuda_get_device();
    const int cc = ggml_cuda_info().devices[id].cc;
@@ -1406,13 +1458,17 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a

    const dim3 block_dims(warp_size, nwarps, 1);

-    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, J, false, prec_src1>), nbytes_shared);
-    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, J,  true, prec_src1>), nbytes_shared);
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, J, /*fallback =*/ false, prec_src1, has_fusion>), nbytes_shared);
+    if constexpr (!has_fusion) {
+        CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, J, /*fallback =*/ true, prec_src1, false>), nbytes_shared);
+    }

-    const int nty  = (args.nrows_x   + config.I - 1) / config.I;
+    const int I_dst = has_fusion ? config.I/2 : config.I; // Rows of dst per tile, see mul_mat_q.
+    const int nty  = (args.nrows_x   + I_dst - 1) / I_dst;
    const int ntx  = (args.ncols_max + config.J - 1) / config.J;
    const int ntzw = args.nchannels_y * args.nsamples_y;
-    const dim3 block_nums_xy_tiling(nty, ntx, ntzw);
+    // Column tiles innermost so that the concurrently running blocks share the same tile of x and read it from L2.
+    const dim3 block_nums_xy_tiling(ntx, nty, ntzw);

    GGML_ASSERT(args.nchannels_y % args.nchannels_x == 0);
    GGML_ASSERT(args.nsamples_y  % args.nsamples_x  == 0);
@@ -1426,9 +1482,10 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
    const uint3 channel_ratio_fd   = init_fastdiv_values(channel_ratio);
    const uint3 sample_ratio_fd    = init_fastdiv_values(sample_ratio);

-    if (!config.stream_k) {
-        mul_mat_q<type, J, fallback, prec_src1><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
-            (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, args.y_scale,
+    if (has_fusion || !config.stream_k) {
+        GGML_ASSERT(nty <= 65535); // gridDim.y limit
+        mul_mat_q<type, J, fallback, prec_src1, has_fusion><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+            (args.x, (const char *) args.fusion.gate, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, args.y_scale, args.fusion.glu_op, args.fusion.glu_limit,
             blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
             channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
             sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
@@ -1458,8 +1515,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
    const dim3 block_nums_fixup(block_nums_stream_k.x, config.I/warp_size, 1);
    const dim3 block_dims_fixup(block_dims.x, block_dims.y/2, block_dims.z);

-    mul_mat_q<type, J, fallback, prec_src1><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
-        (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.y_scale,
+    mul_mat_q<type, J, fallback, prec_src1, has_fusion><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
+        (args.x, (const char *) args.fusion.gate, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.y_scale, args.fusion.glu_op, args.fusion.glu_limit,
         blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
         channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
         sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
@@ -1476,56 +1533,56 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
         ntx_fd);
 }

-template <ggml_type type, bool fallback, ggml_prec prec_src1>
+template <ggml_type type, bool fallback, ggml_prec prec_src1, bool has_fusion>
 void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
    switch (args.J_best) {
        case   8:
-            launch_mul_mat_q<type,   8, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,   8, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  16:
-            launch_mul_mat_q<type,  16, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  16, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  24:
-            launch_mul_mat_q<type,  24, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  24, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  32:
-            launch_mul_mat_q<type,  32, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  32, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  40:
-            launch_mul_mat_q<type,  40, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  40, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  48:
-            launch_mul_mat_q<type,  48, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  48, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  56:
-            launch_mul_mat_q<type,  56, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  56, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  64:
-            launch_mul_mat_q<type,  64, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  64, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  72:
-            launch_mul_mat_q<type,  72, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  72, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  80:
-            launch_mul_mat_q<type,  80, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  80, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  88:
-            launch_mul_mat_q<type,  88, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  88, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case  96:
-            launch_mul_mat_q<type,  96, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type,  96, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case 104:
-            launch_mul_mat_q<type, 104, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type, 104, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case 112:
-            launch_mul_mat_q<type, 112, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type, 112, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case 120:
-            launch_mul_mat_q<type, 120, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type, 120, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        case 128:
-            launch_mul_mat_q<type, 128, fallback, prec_src1>(ctx, args, stream);
+            launch_mul_mat_q<type, 128, fallback, prec_src1, has_fusion>(ctx, args, stream);
            break;
        default:
            fprintf(stderr, "J_best=%d\n", args.J_best);
@@ -1533,14 +1590,26 @@ void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args,
    }
 }

+template <ggml_type type, bool fallback, ggml_prec prec_src1>
+static void mul_mat_q_switch_fusion(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    if constexpr (!fallback && ggml_cuda_mmq_fusion_supported(type)) {
+        if (args.fusion.gate) {
+            mul_mat_q_switch_J<type, fallback, prec_src1, true>(ctx, args, stream);
+            return;
+        }
+    }
+    GGML_ASSERT(!args.fusion.gate);
+    mul_mat_q_switch_J<type, fallback, prec_src1, false>(ctx, args, stream);
+}
+
 template <ggml_type type, ggml_prec prec_src1>
 void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
    if (ggml_cuda_mmq_needs_fallback(args.nrows_x)) {
        constexpr bool fallback = true;
-        mul_mat_q_switch_J<type, fallback, prec_src1>(ctx, args, stream);
+        mul_mat_q_switch_fusion<type, fallback, prec_src1>(ctx, args, stream);
    } else {
        constexpr bool fallback = false;
-        mul_mat_q_switch_J<type, fallback, prec_src1>(ctx, args, stream);
+        mul_mat_q_switch_fusion<type, fallback, prec_src1>(ctx, args, stream);
    }
 }

@@ -1582,6 +1651,9 @@ extern DECL_MMQ_CASE_W4A4(GGML_TYPE_NVFP4);
 // -------------------------------------------------------------------------------------------------------------------------

 void ggml_cuda_mul_mat_q(
-        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
+        const ggml_cuda_mm_fusion_args_host * fusion = nullptr);

 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t n_experts);
+
+bool ggml_cuda_mmq_fusion_is_efficient(ggml_type type, int64_t nrows_x, int64_t ncols, int64_t nchannels, int device);
@@ -196,9 +196,9 @@ static __global__ void rope_neox(const T *            x,
    dst[idst + n_offs/2 + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

-template <bool forward, bool has_ff, typename T>
+template <bool forward, bool has_ff, typename T, typename D>
 static __global__ void rope_multi(const T *            x,
-                                  T *                  dst,
+                                  D *                  dst,
                                  const int            ne00,
                                  const int            ne01,
                                  const int            ne02,
@@ -219,6 +219,8 @@ static __global__ void rope_multi(const T *            x,
                                  const float *        freq_factors,
                                  const mrope_sections sections,
                                  const bool           is_imrope,
+                                  const int64_t *      row_indices,
+                                  const int            set_rows_stride,
                                  const bool           inplace) {
    const int i0 = 2 * (blockDim.y * blockIdx.y + threadIdx.y);

@@ -236,12 +238,20 @@ static __global__ void rope_multi(const T *            x,
    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

    ggml_cuda_pdl_sync();
+
+    // Fusion optimization: ROPE + VIEW + SET_ROWS.
+    // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
+    if (set_rows_stride != 0) {
+        idst = i1 * s1 + i0 / 2;
+        idst += row_indices[i2] * set_rows_stride;
+    }
+
    if (i0 < n_offs || i0 >= n_offs + n_dims) {
        if (inplace) {
            return;
        }
-        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
-        dst[idst + i0/2 + 1] = x[ix + i0/2 + 1];
+        dst[idst + i0/2 + 0] = ggml_cuda_cast<D>(x[ix + i0/2 + 0]);
+        dst[idst + i0/2 + 1] = ggml_cuda_cast<D>(x[ix + i0/2 + 1]);

        return;
    }
@@ -286,8 +296,8 @@ static __global__ void rope_multi(const T *            x,
    const float x0 = x[ix + n_offs/2 + 0];
    const float x1 = x[ix + n_offs/2 + n_dims/2];

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

 template <bool forward, bool has_ff, typename T>
@@ -443,9 +453,9 @@ static void rope_neox_cuda(const T *            x,
    }
 }

-template <bool forward, typename T>
+template <bool forward, typename T, typename D>
 static void rope_multi_cuda(const T *            x,
-                            T *                  dst,
+                            D *                  dst,
                            const int            ne00,
                            const int            ne01,
                            const int            ne02,
@@ -467,6 +477,8 @@ static void rope_multi_cuda(const T *            x,
                            const float *        freq_factors,
                            const mrope_sections sections,
                            const bool           is_imrope,
+                            const int64_t *      row_indices,
+                            const int            set_rows_stride,
                            const bool           inplace,
                            cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
@@ -478,14 +490,14 @@ static void rope_multi_cuda(const T *            x,

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

@@ -648,14 +660,21 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
            GGML_ABORT("fatal error");
        }
    } else if (is_mrope && !is_vision) {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_multi_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
-                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_multi_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
-                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
+        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
+            rope_multi_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                   s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
+                                                   ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope,
+                                                   row_indices, set_rows_stride, inplace, stream);
+        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
+            rope_multi_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                  s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
+                                                  ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope,
+                                                  row_indices, set_rows_stride, inplace, stream);
+        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
+            rope_multi_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                 s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
+                                                 ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope,
+                                                 row_indices, set_rows_stride, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -151,9 +151,10 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const floa
        case 3:  launch_kernel(std::integral_constant<int, 3 >{}); break;
        case 4:  launch_kernel(std::integral_constant<int, 4 >{}); break;
        case 5:  launch_kernel(std::integral_constant<int, 5 >{}); break;
+        case 7:  launch_kernel(std::integral_constant<int, 7 >{}); break;
        case 9:  launch_kernel(std::integral_constant<int, 9 >{}); break;
        case 15: launch_kernel(std::integral_constant<int, 15>{}); break;
-        default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 9, 15 right now.");
+        default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 7, 9, 15 right now.");
    }
 }

@@ -0,0 +1,141 @@
+#include "top-k-mask.cuh"
+#include "convert.cuh"
+
+#define CUDA_TOP_K_MASK_BLOCK_SIZE 512
+
+// one block per row: fill the row with -inf, then copy the mask cells of the live picks and the extra cells
+template <typename T>
+static __global__ void top_k_mask_kernel(
+        const T * mask, const int32_t * top_k, const float * score, const int32_t * blocks, const int32_t * extra, T * dst,
+        const int64_t n_kv, const int64_t ne1, const int64_t ne2,
+        const int64_t m1, const int64_t m2, const int64_t m3,
+        const int64_t n_top, const int64_t n_pick, const int64_t n_blk, const int64_t n_extra,
+        const int64_t sm1, const int64_t sm2, const int64_t sm3,
+        const int64_t sd1, const int64_t sd2, const int64_t sd3,
+        const int64_t sk0, const int64_t sk1, const int64_t sk2, const int64_t sk3,
+        const int64_t ss0, const int64_t ss1, const int64_t ss2, const int64_t ss3,
+        const int64_t sb0, const int64_t sb1,
+        const int64_t se0, const int64_t se1, const int64_t se2, const int64_t se3,
+        const bool vec) {
+    ggml_cuda_pdl_lc();
+
+    const int64_t ir = blockIdx.x;
+    const int64_t i1 = ir % ne1;
+    const int64_t i2 = (ir / ne1) % ne2;
+    const int64_t i3 = ir / (ne1*ne2);
+
+    const T       * mask_row  = mask  + (i1 % m1)*sm1 + (i2 % m2)*sm2 + (i3 % m3)*sm3;
+    T             * dst_row   = dst   + i1*sd1 + i2*sd2 + i3*sd3;
+    const int32_t * top_k_row = top_k + i1*sk1 + i2*sk2 + i3*sk3;
+
+    ggml_cuda_pdl_sync();
+
+    const T ninf = ggml_cuda_cast<T>(-INFINITY);
+    if (vec) {
+        constexpr int nv = sizeof(int4)/sizeof(T);
+        T fill[nv];
+#pragma unroll
+        for (int j = 0; j < nv; ++j) {
+            fill[j] = ninf;
+        }
+        const int4 fill4 = *(const int4 *) fill;
+        for (int64_t i = threadIdx.x; i < n_kv/nv; i += blockDim.x) {
+            ((int4 *) dst_row)[i] = fill4;
+        }
+    } else {
+        for (int64_t i = threadIdx.x; i < n_kv; i += blockDim.x) {
+            dst_row[i] = ninf;
+        }
+    }
+
+    __syncthreads();
+
+    // one thread per pick or extra cell; cells kept twice get the same value
+    for (int64_t i = threadIdx.x; i < n_top + n_extra; i += blockDim.x) {
+        if (i >= n_top) {
+            const int32_t c = extra[(i - n_top)*se0 + i1*se1 + i2*se2 + i3*se3];
+            if (c >= 0 && c < n_kv) {
+                dst_row[c] = mask_row[c];
+            }
+            continue;
+        }
+
+        const int32_t p = top_k_row[i*sk0];
+        if (p < 0 || p >= n_pick) {
+            continue;
+        }
+        if (score && !(score[p*ss0 + i1*ss1 + i2*ss2 + i3*ss3] > -INFINITY)) {
+            continue;
+        }
+        if (!blocks) {
+            if (p < n_kv) {
+                dst_row[p] = mask_row[p];
+            }
+            continue;
+        }
+        for (int64_t k = 0; k < n_blk; ++k) {
+            const int32_t c = blocks[k*sb0 + p*sb1];
+            if (c >= 0 && c < n_kv) {
+                dst_row[c] = mask_row[c];
+            }
+        }
+    }
+}
+
+template <typename T>
+static void top_k_mask_cuda(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * mask   = dst->src[0];
+    const ggml_tensor * top_k  = dst->src[1];
+    const ggml_tensor * score  = dst->src[2];
+    const ggml_tensor * blocks = dst->src[3];
+    const ggml_tensor * extra  = dst->src[4];
+
+    const int64_t n_kv = mask->ne[0];
+
+    GGML_ASSERT(dst->nb[0] == sizeof(T) && mask->nb[0] == sizeof(T));
+    GGML_ASSERT(dst->nb[1] % sizeof(T) == 0 && mask->nb[1] % sizeof(T) == 0);
+
+    // 16-byte stores need every dst row aligned
+    const bool vec = (n_kv*sizeof(T)) % sizeof(int4) == 0 &&
+        dst->nb[1] % sizeof(int4) == 0 && dst->nb[2] % sizeof(int4) == 0 && dst->nb[3] % sizeof(int4) == 0 &&
+        (uintptr_t) dst->data % sizeof(int4) == 0;
+
+    const int64_t n_pick = blocks ? blocks->ne[1] : score ? score->ne[0] : n_kv;
+
+    const size_t si = sizeof(int32_t);
+    const size_t sf = sizeof(float);
+
+    const dim3 block_dims(CUDA_TOP_K_MASK_BLOCK_SIZE, 1, 1);
+    const dim3 grid_dims(ggml_nrows(dst), 1, 1);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(grid_dims, block_dims, 0, ctx.stream());
+
+    ggml_cuda_kernel_launch(top_k_mask_kernel<T>, launch_params,
+            (const T *) mask->data, (const int32_t *) top_k->data,
+            score  ? (const float   *) score->data  : nullptr,
+            blocks ? (const int32_t *) blocks->data : nullptr,
+            extra  ? (const int32_t *) extra->data  : nullptr,
+            (T *) dst->data,
+            n_kv, dst->ne[1], dst->ne[2],
+            mask->ne[1], mask->ne[2], mask->ne[3],
+            top_k->ne[0], n_pick, blocks ? blocks->ne[0] : 1, extra ? extra->ne[0] : 0,
+            mask->nb[1]/sizeof(T), mask->nb[2]/sizeof(T), mask->nb[3]/sizeof(T),
+            dst->nb[1]/sizeof(T),  dst->nb[2]/sizeof(T),  dst->nb[3]/sizeof(T),
+            top_k->nb[0]/si, top_k->nb[1]/si, top_k->nb[2]/si, top_k->nb[3]/si,
+            score  ? score->nb[0]/sf  : 0, score ? score->nb[1]/sf : 0, score ? score->nb[2]/sf : 0, score ? score->nb[3]/sf : 0,
+            blocks ? blocks->nb[0]/si : 0, blocks ? blocks->nb[1]/si : 0,
+            extra  ? extra->nb[0]/si  : 0, extra ? extra->nb[1]/si : 0, extra ? extra->nb[2]/si : 0, extra ? extra->nb[3]/si : 0,
+            vec);
+}
+
+void ggml_cuda_op_top_k_mask(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    switch (dst->type) {
+        case GGML_TYPE_F16:
+            top_k_mask_cuda<half>(ctx, dst);
+            break;
+        case GGML_TYPE_F32:
+            top_k_mask_cuda<float>(ctx, dst);
+            break;
+        default:
+            GGML_ABORT("unsupported type");
+    }
+}
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_top_k_mask(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
@@ -76,7 +76,7 @@ __device__ void sqrt_softplus_warp_inplace(float (&vals)[experts_per_thread], co
    for (int i = 0; i < experts_per_thread; i++) {
        const int  idx    = lane + i * WARP_SIZE;
        const bool active = !use_limit || (idx < limit);
-        vals[i]           = active ? sqrtf(vals[i] > 20.0f ? vals[i] : logf(1.0f + expf(vals[i]))) : -INFINITY;
+        vals[i]           = active ? sqrtf(vals[i] > 20.0f ? vals[i] : log1pf(expf(vals[i]))) : -INFINITY;
    }
 }

@@ -91,7 +91,7 @@ static __device__ __forceinline__ float op_expm1(float x) {
 }

 static __device__ __forceinline__ float op_softplus(float x) {
-    return (x > 20.0f) ? x : logf(1.0f + expf(x));
+    return (x > 20.0f) ? x : log1pf(expf(x));
 }

 static __device__ __forceinline__ float op_elu(float x) {
@@ -121,3 +121,18 @@ __device__ __forceinline__ float ggml_cuda_op_swiglu_clamp_single(float gate, fl

    return ggml_cuda_op_silu_single(gate) * up;
 }
+
+__device__ __forceinline__ float ggml_cuda_op_glu_single(ggml_glu_op op, float gate, float up, float limit) {
+    switch (op) {
+        case GGML_GLU_OP_SWIGLU:
+            return up * ggml_cuda_op_silu_single(gate);
+        case GGML_GLU_OP_GEGLU:
+            return up * ggml_cuda_op_gelu_single(gate);
+        case GGML_GLU_OP_SWIGLU_OAI:
+            return ggml_cuda_op_swiglu_oai_single(gate, up);
+        case GGML_GLU_OP_SWIGLU_CLAMP:
+            return ggml_cuda_op_swiglu_clamp_single(gate, up, limit);
+        default:
+            return up * gate;
+    }
+}
@@ -40,18 +40,18 @@ static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4, con
    const uint32_t q_odd  = (q4 >> 4);

    // Perform lookups in the lower half of the table (indices 0-7).
-    uint32_t v_even_low = __builtin_amdgcn_perm(values[1], values[0], q_even & 0x07070707);
-    uint32_t v_odd_low = __builtin_amdgcn_perm(values[1], values[0], q_odd & 0x07070707);
+    uint32_t v_even_low = ggml_hip_perm(values[1], values[0], q_even & 0x07070707);
+    uint32_t v_odd_low  = ggml_hip_perm(values[1], values[0], q_odd & 0x07070707);

    // Perform lookups in the upper half of the table (indices 8-15).
-    uint32_t v_even_high = __builtin_amdgcn_perm(values[3], values[2], q_even & 0x07070707);
-    uint32_t v_odd_high = __builtin_amdgcn_perm(values[3], values[2], q_odd & 0x07070707);
+    uint32_t v_even_high = ggml_hip_perm(values[3], values[2], q_even & 0x07070707);
+    uint32_t v_odd_high  = ggml_hip_perm(values[3], values[2], q_odd & 0x07070707);

    // Select between the low and high results based on the MSB of each index nibble.
    uint32_t mask_even = 0x03020100 | ((q_even & 0x08080808) >> 1);
-    uint32_t res_x = __builtin_amdgcn_perm(v_even_high, v_even_low, mask_even);
+    uint32_t res_x     = ggml_hip_perm(v_even_high, v_even_low, mask_even);
    uint32_t mask_odd = 0x03020100 | ((q_odd & 0x08080808) >> 1);
-    uint32_t res_y = __builtin_amdgcn_perm(v_odd_high, v_odd_low, mask_odd);
+    uint32_t res_y     = ggml_hip_perm(v_odd_high, v_odd_low, mask_odd);

    return make_int2(res_x, res_y);
 #elif !defined(GGML_USE_MUSA)
@@ -720,10 +720,10 @@ static __device__ __forceinline__ float vec_dot_q1_0_q8_1(
        const uint32_t q2_indices = (q2_bits & 0x01) | ((q2_bits & 0x02) << 7) | ((q2_bits & 0x04) << 14) | ((q2_bits & 0x08) << 21);
        const uint32_t q3_indices = (q3_bits & 0x01) | ((q3_bits & 0x02) << 7) | ((q3_bits & 0x04) << 14) | ((q3_bits & 0x08) << 21);

-        const int v0 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q0_indices);
-        const int v1 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q1_indices);
-        const int v2 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q2_indices);
-        const int v3 = __builtin_amdgcn_perm(0x000001FF, 0x000001FF, q3_indices);
+        const int v0 = ggml_hip_perm(0x000001FF, 0x000001FF, q0_indices);
+        const int v1 = ggml_hip_perm(0x000001FF, 0x000001FF, q1_indices);
+        const int v2 = ggml_hip_perm(0x000001FF, 0x000001FF, q2_indices);
+        const int v3 = ggml_hip_perm(0x000001FF, 0x000001FF, q3_indices);
 #else
        // unpack crumbs into nibble indices
        const int n0 = __byte_perm(0x11100100, 0x11100100, q >> 0); // [0, 1, 4, 5] [ 8,  9, 12, 13]
@@ -777,8 +777,8 @@ static __device__ __forceinline__ float vec_dot_q2_0_q8_1(
        const uint32_t qx_indices = (q & 0x03) | ((q & 0x0C) << 6) | ((q & 0x30) << 12) | ((q & 0xC0) << 18);
        const uint32_t qy_bits    = q >> 8;
        const uint32_t qy_indices = (qy_bits & 0x03) | ((qy_bits & 0x0C) << 6) | ((qy_bits & 0x30) << 12) | ((qy_bits & 0xC0) << 18);
-        const int qx = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qx_indices);
-        const int qy = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qy_indices);
+        const int      qx         = ggml_hip_perm(0x020100FF, 0x020100FF, qx_indices);
+        const int      qy         = ggml_hip_perm(0x020100FF, 0x020100FF, qy_indices);
 #else
        // unpack even and odd crumbs into byte values
        const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
@@ -311,3 +311,22 @@ static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigne

    return (ne_any_bit >> 7) * 0xff;
 }
+
+// v_perm_b32 with selector bytes in [0, 7]: 0-3 pick bytes of lo, 4-7 pick bytes of hi.
+static __device__ __forceinline__ uint32_t ggml_hip_perm(const uint32_t hi, const uint32_t lo, const uint32_t sel) {
+#if __has_builtin(__builtin_amdgcn_is_invocable)
+    if (__builtin_amdgcn_is_invocable(__builtin_amdgcn_perm)) {
+        return __builtin_amdgcn_perm(hi, lo, sel);
+    }
+    const uint64_t src = ((uint64_t) hi << 32) | lo;
+    uint32_t res = 0;
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t byte_sel = (sel >> (8 * i)) & 0x07;
+        res |= (uint32_t) ((src >> (8 * byte_sel)) & 0xFF) << (8 * i);
+    }
+    return res;
+#else
+    return __builtin_amdgcn_perm(hi, lo, sel);
+#endif // __has_builtin(__builtin_amdgcn_is_invocable)
+}
@@ -22,6 +22,10 @@
 #define HWCAP2_I8MM (1ULL << 13)
 #endif

+#if !defined(HWCAP2_SVEI8MM)
+#define HWCAP2_SVEI8MM (1ULL << 9)
+#endif
+
 #if !defined(HWCAP_ASIMDDP)
 #define HWCAP_ASIMDDP (1 << 20)
 #endif
@@ -63,6 +67,10 @@
 #define PF_ARM_SVE2_INSTRUCTIONS_AVAILABLE 47
 #endif

+#if !defined(PF_ARM_SVE_I8MM_INSTRUCTIONS_AVAILABLE)
+#define PF_ARM_SVE_I8MM_INSTRUCTIONS_AVAILABLE 57
+#endif
+
 #if !defined(PF_ARM_V82_I8MM_INSTRUCTIONS_AVAILABLE)
 #define PF_ARM_V82_I8MM_INSTRUCTIONS_AVAILABLE 66
 #endif
@@ -87,6 +95,7 @@ typedef struct ggml_feats_arch64_runtime {
    bool has_sve;
    bool has_sve2;
    bool has_i8mm;
+    bool has_svei8mm;
    bool has_sme;
    bool has_sme2;
    int sve_cnt;
@@ -104,6 +113,7 @@ static inline ggml_feats_arch64_runtime_t ggml_feats_get_arch64_runtime(void) {
    runtime_feat.has_sve     = !!(hwcap & HWCAP_SVE);
    runtime_feat.has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
    runtime_feat.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+    runtime_feat.has_svei8mm = !!(hwcap2 & HWCAP2_SVEI8MM);
    runtime_feat.has_sme     = !!(hwcap2 & HWCAP2_SME);
    runtime_feat.has_sme2    = !!(hwcap2 & HWCAP2_SME2);

@@ -153,6 +163,7 @@ static inline ggml_feats_arch64_runtime_t ggml_feats_get_arch64_runtime(void) {
    runtime_feat.has_sve     = IsProcessorFeaturePresent(PF_ARM_SVE_INSTRUCTIONS_AVAILABLE) != 0;
    runtime_feat.has_sve2    = IsProcessorFeaturePresent(PF_ARM_SVE2_INSTRUCTIONS_AVAILABLE) != 0;
    runtime_feat.has_i8mm    = IsProcessorFeaturePresent(PF_ARM_V82_I8MM_INSTRUCTIONS_AVAILABLE) != 0;
+    runtime_feat.has_svei8mm = IsProcessorFeaturePresent(PF_ARM_SVE_I8MM_INSTRUCTIONS_AVAILABLE) != 0;
    runtime_feat.has_sme     = IsProcessorFeaturePresent(PF_ARM_SME_INSTRUCTIONS_AVAILABLE) != 0;
    runtime_feat.has_sme2    = IsProcessorFeaturePresent(PF_ARM_SME2_INSTRUCTIONS_AVAILABLE) != 0;

@@ -51,6 +51,13 @@ endif()
 set(TARGET_NAME ggml-hexagon)
 ggml_add_backend_library(${TARGET_NAME}
    ggml-hexagon.cpp
+    htp-common.h
+    htp-kparams.cpp
+    htp-kparams.h
+    htp-opnode.cpp
+    htp-opnode.h
+    htp-fusion.cpp
+    htp-fusion.h
    htp-repack.cpp
    htp-repack.h
    htp-drv.cpp
@@ -0,0 +1,106 @@
+#ifndef HTP_COMMON_H
+#define HTP_COMMON_H
+
+#include "htp-opnode.h"
+#include "htp-repack.h"
+
+struct htp_tensor_extra {
+    std::vector<uint8_t> shadow_buf;
+    size_t               shadow_size { 0 };
+    uint32_t             flags { 0 };
+};
+
+extern int    opt_arch;
+extern size_t opt_ndev;
+extern size_t opt_nhvx;
+extern int    opt_nhmx;
+extern size_t opt_vmem;
+extern int    opt_etm;
+extern int    opt_verbose;
+extern int    opt_profile;
+extern bool   opt_hostbuf;
+extern bool   opt_dma64;
+extern int    opt_power_mode;
+
+extern size_t opt_mbuf_dyn;
+extern size_t opt_mbuf_static;
+extern size_t opt_mbuf_total;
+
+extern int    opt_mm_select;
+extern int    opt_fa_select;
+extern int    opt_fa_split;
+extern int    opt_gdn_select;
+extern int    opt_ar_select;
+extern int    opt_ar_scatter;
+
+extern int    opt_opbatch;
+extern int    opt_opqueue;
+extern int    opt_optrace;
+extern int    opt_oppoll;
+extern int    opt_opfusion;
+extern int    opt_rt_repack;
+
+#ifndef HEX_VERBOSE
+#define HEX_VERBOSE(...) \
+    if (opt_verbose) GGML_LOG_DEBUG(__VA_ARGS__)
+#endif
+
+static inline bool htp_tensor_is_fuseable(const struct ggml_tensor * t) {
+    if (!t || !t->extra) return false;
+    const auto * extra = (const struct htp_tensor_extra *) t->extra;
+    return (extra->flags & HTP_TENSOR_FUSEABLE) != 0;
+}
+
+static inline bool htp_can_rt_repack(const struct ggml_tensor * t) {
+    if (opt_rt_repack == 0 || !htp_is_tiled_weight_tensor(t)) {
+        return false;
+    }
+    if (t->buffer && ggml_backend_buft_is_host(t->buffer->buft)) {
+        return false;
+    }
+    return opt_rt_repack > 1 || ggml_is_quantized(t->type);
+}
+
+static inline bool htp_tensor_is_tiled(const struct ggml_tensor * t) {
+    const auto * extra = (const struct htp_tensor_extra *) t->extra;
+    if (extra && (extra->flags & HTP_TENSOR_TILED)) {
+        return true;
+    }
+    return htp_can_rt_repack(t);
+}
+
+static inline bool htp_needs_repack(const struct ggml_tensor * t) {
+    const auto * extra = (const struct htp_tensor_extra *) t->extra;
+    if (extra && (extra->flags & HTP_TENSOR_TILED)) {
+        return false;
+    }
+    return htp_can_rt_repack(t);
+}
+
+static inline bool htp_tensors_overlap(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    const uintptr_t a0 = (uintptr_t) a->data;
+    const uintptr_t b0 = (uintptr_t) b->data;
+    const uintptr_t a1 = a0 + ggml_nbytes(a);
+    const uintptr_t b1 = b0 + ggml_nbytes(b);
+
+    return a0 < b1 && b0 < a1;
+}
+
+static inline bool htp_can_row_partition(const struct ggml_tensor * t) {
+    if (t->ne[1] > 1 && (t->nb[1] & 127) != 0) return false;
+    if (t->ne[2] > 1 && (t->nb[2] & 127) != 0) return false;
+    if (t->ne[3] > 1 && (t->nb[3] & 127) != 0) return false;
+    return true;
+}
+
+struct ggml_hexagon_session;
+
+const char * htp_session_name(const struct ggml_hexagon_session * sess);
+size_t       htp_session_vtcm_size(const struct ggml_hexagon_session * sess);
+uint32_t     htp_session_n_threads(const struct ggml_hexagon_session * sess);
+uint32_t     htp_session_n_hmx(const struct ggml_hexagon_session * sess);
+uint32_t     htp_session_mdev_count(const struct ggml_hexagon_session * sess);
+
+htp_op_code  htp_op_remap(const struct ggml_tensor * t);
+
+#endif // HTP_COMMON_H
@@ -0,0 +1,649 @@
+#include "htp-fusion.h"
+#include "htp-kparams.h"
+#include "ggml-impl.h"
+
+#include <algorithm>
+#include <cstring>
+
+static inline bool htp_same_tensor_data(const ggml_tensor * a, const ggml_tensor * b) {
+    return a && b && (a == b || a->data == b->data);
+}
+
+static inline bool htp_same_shape_and_contiguity(const ggml_tensor * a, const ggml_tensor * b) {
+    return ggml_are_same_shape(a, b) && (ggml_is_contiguous(a) == ggml_is_contiguous(b));
+}
+
+static inline bool htp_inplace_overlap_unsafe(const ggml_tensor * dst, const ggml_tensor * in) {
+    return htp_tensors_overlap(dst, in) && (dst->data != in->data || !ggml_are_same_stride(dst, in));
+}
+
+static inline bool htp_inputs_overlap_unsafe(const htp_opnode & prev, const ggml_tensor * dst, const ggml_tensor * extra = nullptr) {
+    if (extra && htp_inplace_overlap_unsafe(dst, extra)) return true;
+    for (const auto * in : prev.inputs) {
+        if (htp_inplace_overlap_unsafe(dst, in)) return true;
+    }
+    return false;
+}
+
+static inline bool htp_match_consumer(
+    const htp_opnode & prev,
+    const htp_opnode & cur,
+    bool commutative,
+    const ggml_tensor ** other_src = nullptr
+) {
+    if (prev.outputs.size() != 1) {
+        return false;
+    }
+    const ggml_tensor * prev_out = prev.dst();
+    if (!htp_tensor_is_fuseable(prev_out)) {
+        return false;
+    }
+    if (htp_same_tensor_data(cur.src0(), prev_out)) {
+        if (other_src) *other_src = cur.src1();
+        return true;
+    }
+    if (commutative && htp_same_tensor_data(cur.src1(), prev_out)) {
+        if (other_src) *other_src = cur.src0();
+        return true;
+    }
+    return false;
+}
+
+static bool is_supported_mul_mat_nx_kernel(const ggml_tensor * src0, const struct htp_mm_kernel_params * kparams) {
+    if (kparams->n_hmx) {
+        return kparams->kernel_type == HTP_MM_KERNEL_HMX_2D;
+    }
+    return htp_tensor_is_tiled(src0) && (kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_ROW);
+}
+
+static bool is_supported_mul_mat_id_nx_kernel(const ggml_tensor * src0, const struct htp_mm_kernel_params * kparams) {
+    if (kparams->n_hmx) {
+        return kparams->kernel_type == HTP_MM_KERNEL_HMX_2D;
+    }
+    return htp_tensor_is_tiled(src0) &&
+           (kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_ROW || kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_BLOCK);
+}
+
+// --- Fusion rules ---
+
+static bool fuse_allreduce_add(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (opt_ar_select != 2 || prev.opcode != HTP_OP_ALLREDUCE) return false;
+
+    const auto * ar_kparams = prev.kparams<htp_allreduce_kernel_params>();
+    const uint32_t rank    = (uint32_t) ar_kparams->rank;
+    const uint32_t n_ranks = (uint32_t) ar_kparams->n_ranks;
+
+    const ggml_tensor * ar_local = prev.inputs[rank];
+    if (!htp_tensor_is_fuseable(ar_local)) return false;
+
+    const ggml_tensor * add_dst    = cur.dst();
+    const ggml_tensor * res_tensor = nullptr;
+    if (htp_same_tensor_data(cur.src0(), ar_local)) {
+        res_tensor = cur.src1();
+    } else if (htp_same_tensor_data(cur.src1(), ar_local)) {
+        res_tensor = cur.src0();
+    } else {
+        return false;
+    }
+
+    if (ar_local->type != res_tensor->type) return false;
+
+    const bool is_same_shape = ggml_are_same_shape(ar_local, res_tensor);
+    const bool is_row_bcast  = !is_same_shape && (ar_local->ne[0] == res_tensor->ne[0] &&
+                                                  res_tensor->ne[1] == 1 && res_tensor->ne[2] == 1 && res_tensor->ne[3] == 1);
+    if (!is_same_shape && !is_row_bcast) return false;
+
+    if (is_same_shape && !ggml_are_same_stride(ar_local, res_tensor)) {
+        return false;
+    }
+    if (ggml_is_contiguous(ar_local) != ggml_is_contiguous(add_dst)) return false;
+
+    const bool is_shard_ok = n_ranks <= HTP_OP_MAX_OUTPUTS && add_dst->data == ar_local->data;
+
+    struct htp_allreduce_kernel_params kparams;
+    if (!htp_kparams_solve_allreduce(sess, add_dst, rank, n_ranks, true, is_row_bcast, is_shard_ok, &kparams)) {
+        HEX_VERBOSE("ggml-hex: %s skip ALLREDUCE_ADD fusion: solver failed\n", htp_session_name(sess));
+        return false;
+    }
+
+    const bool scatter_ok = kparams.n_dsts > 1;
+    kparams.mode = scatter_ok ? HTP_ALLREDUCE_SHARDED_FANOUT : HTP_ALLREDUCE_FULL;
+
+    for (uint32_t r = 0; r < n_ranks; r++) {
+        const ggml_tensor * ar_src = prev.inputs[r];
+        if (!htp_tensors_overlap(add_dst, ar_src)) continue;
+        if (scatter_ok && r == rank) continue;
+        HEX_VERBOSE("ggml-hex: %s skip ALLREDUCE_ADD fusion: dst overlaps allreduce src %u\n", htp_session_name(sess), r);
+        return false;
+    }
+
+    prev.opcode = HTP_OP_ALLREDUCE_ADD;
+    prev.inputs.push_back(res_tensor);
+    if (kparams.mode != HTP_ALLREDUCE_SHARDED_FANOUT) {
+        prev.outputs = { add_dst };
+    }
+    prev.fused.push_back(cur.node);
+    prev.set_kparams(kparams);
+    return true;
+}
+
+static bool fuse_rms_norm_mul(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (prev.opcode != HTP_OP_RMS_NORM) return false;
+
+    const ggml_tensor * weight = nullptr;
+    if (!htp_match_consumer(prev, cur, /*commutative=*/ true, &weight)) return false;
+
+    const ggml_tensor * src0 = prev.src0();
+    const ggml_tensor * dst  = cur.dst();
+    if (src0->type != GGML_TYPE_F32 || src0->ne[0] != weight->ne[0] || !htp_same_shape_and_contiguity(prev.dst(), dst)) return false;
+    if (htp_inputs_overlap_unsafe(prev, dst, weight)) return false;
+
+    const bool is_row_bcast  = (weight->ne[1] == 1 && weight->ne[2] == 1 && weight->ne[3] == 1);
+    const bool is_same_shape = ggml_are_same_shape(src0, weight);
+    if (!is_row_bcast && !is_same_shape) return false;
+
+    struct htp_unary_kernel_params kparams;
+    if (!htp_kparams_solve_rms_norm_mul(sess, src0, weight, dst, &kparams)) {
+        HEX_VERBOSE("ggml-hex: %s skip RMS_NORM_MUL fusion: VTCM needed (%d) > budget (%zu)\n",
+                    htp_session_name(sess), kparams.vtcm_size, htp_session_vtcm_size(sess));
+        return false;
+    }
+
+    prev.opcode  = HTP_OP_RMS_NORM_MUL;
+    prev.inputs  = { src0, weight };
+    prev.outputs = { dst };
+    prev.fused.push_back(cur.node);
+    prev.set_kparams(kparams);
+    return true;
+}
+
+static bool fuse_rms_norm_scale(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    GGML_UNUSED(sess);
+    if (prev.opcode != HTP_OP_RMS_NORM && prev.opcode != HTP_OP_RMS_NORM_MUL) return false;
+    if (!htp_match_consumer(prev, cur, /*commutative=*/ false)) return false;
+
+    const ggml_tensor * src0 = prev.src0();
+    const ggml_tensor * dst  = cur.dst();
+    if (src0->type != GGML_TYPE_F32 || !htp_same_shape_and_contiguity(prev.dst(), dst)) return false;
+    if (htp_inputs_overlap_unsafe(prev, dst)) return false;
+
+    float scale = 0.0f;
+    float bias  = 0.0f;
+    memcpy(&scale, (const float *) cur.params + 0, sizeof(float));
+    memcpy(&bias,  (const float *) cur.params + 1, sizeof(float));
+    if (bias != 0.0f || scale == 0.0f) return false;
+
+    float prev_scale = 0.0f;
+    memcpy(&prev_scale, &prev.params[1], sizeof(float));
+    if (prev_scale != 0.0f) {
+        scale *= prev_scale;
+    }
+    memcpy(&prev.params[1], &scale, sizeof(float));
+
+    prev.outputs[0] = dst;
+    prev.fused.push_back(cur.node);
+    return true;
+}
+
+static bool fuse_mul_mat_add(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (prev.opcode != HTP_OP_MUL_MAT) return false;
+
+    const ggml_tensor * src2 = nullptr;
+    if (!htp_match_consumer(prev, cur, /*commutative=*/ true, &src2)) return false;
+    if (src2->type != GGML_TYPE_F32) return false;
+
+    const ggml_tensor * src0 = prev.node->src[0];
+    const ggml_tensor * src1 = prev.node->src[1];
+    const ggml_tensor * dst  = cur.node;
+    if (!ggml_are_same_shape(prev.node, dst)) return false;
+
+    const bool prev_collapsed = htp_matmul_can_collapse(src0, src1, prev.node);
+    const bool cur_collapsed  = htp_matmul_add_can_collapse(src0, src1, src2, dst);
+    if (prev_collapsed != cur_collapsed) {
+        return false;
+    }
+
+    const auto * orig_kparams = prev.kparams<htp_mm_kernel_params>();
+    struct htp_mm_kernel_params kparams;
+    if (!htp_kparams_solve_matmul_add(sess, src0, src1, src2, dst, &kparams)) {
+        return false;
+    }
+
+    const int src1_nrows = src1->ne[1] * src1->ne[2] * src1->ne[3];
+    if (kparams.n_hmx == 0 && src1_nrows != 1) return false;
+
+    if (kparams.n_hmx > 0 && orig_kparams->n_hmx > 0) {
+        if (kparams.m_chunk < orig_kparams->m_chunk ||
+            kparams.n_chunk < orig_kparams->n_chunk ||
+            kparams.n_act_threads < orig_kparams->n_act_threads) {
+            HEX_VERBOSE("ggml-hex: %s skip MUL_MAT_ADD fusion: HMX efficiency reduced (m %d->%d, n %d->%d, th %d->%d)\n",
+                        htp_session_name(sess), orig_kparams->m_chunk, kparams.m_chunk,
+                        orig_kparams->n_chunk, kparams.n_chunk,
+                        orig_kparams->n_act_threads, kparams.n_act_threads);
+            return false;
+        }
+    }
+
+    prev.opcode  = HTP_OP_MUL_MAT_ADD;
+    prev.inputs  = { src0, src1, src2 };
+    prev.outputs = { dst };
+    htp_opnode_collapse(sess, prev);
+    prev.fused.push_back(cur.node);
+    prev.set_kparams(kparams);
+    return true;
+}
+
+static bool fuse_mul_mat_nx(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (prev.opcode != HTP_OP_MUL_MAT && prev.opcode != HTP_OP_MUL_MAT_NX) return false;
+
+    const uint32_t curr_n = (uint32_t) prev.outputs.size();
+    if (curr_n >= HTP_OP_MAX_OUTPUTS || curr_n + 1 >= HTP_OP_MAX_INPUTS) return false;
+
+    const ggml_tensor * w0     = prev.inputs[0];
+    const ggml_tensor * x_orig = prev.node->src[1];
+    const ggml_tensor * w_in   = cur.node->src[0];
+    const ggml_tensor * x_in   = cur.node->src[1];
+    const ggml_tensor * d_in   = cur.node;
+
+    auto * orig_kparams = prev.kparams<htp_mm_kernel_params>();
+    const auto * cur_kparams = cur.kparams<htp_mm_kernel_params>();
+
+    if (x_in != x_orig || w_in->type != w0->type || w_in->ne[0] != w0->ne[0] ||
+        w0->ne[2] != 1 || w0->ne[3] != 1 || w_in->ne[2] != 1 || w_in->ne[3] != 1 ||
+        htp_tensor_is_tiled(w_in) != htp_tensor_is_tiled(w0) ||
+        orig_kparams->n_hmx != cur_kparams->n_hmx) {
+        return false;
+    }
+
+    const bool collapse = htp_matmul_can_collapse(w0, x_orig, prev.node) && htp_matmul_can_collapse(w_in, x_in, d_in);
+    if ((x_orig->ne[2] > 1 || x_orig->ne[3] > 1) && !collapse) return false;
+
+    if (prev.opcode == HTP_OP_MUL_MAT_NX) {
+        orig_kparams->n_weights = curr_n + 1;
+    } else {
+        struct htp_mm_kernel_params kparams;
+        const ggml_tensor x_collapsed = collapse ? htp_tensor_collapse_rows(x_orig) : *x_orig;
+        if (!htp_kparams_solve_mmnx(sess, w0, &x_collapsed, curr_n + 1, &kparams)) return false;
+        if (!is_supported_mul_mat_nx_kernel(w0, &kparams)) return false;
+
+        prev.set_kparams(kparams);
+        prev.opcode = HTP_OP_MUL_MAT_NX;
+    }
+
+    prev.inputs.insert(prev.inputs.begin() + curr_n, w_in);
+    prev.outputs.push_back(collapse ? prev.add_dummy(htp_tensor_collapse_rows(d_in)) : d_in);
+    prev.fused.push_back(cur.node);
+    return true;
+}
+
+static bool fuse_mul_mat_id_nx(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (prev.opcode != HTP_OP_MUL_MAT_ID && prev.opcode != HTP_OP_MUL_MAT_ID_NX) return false;
+
+    const uint32_t curr_n = (uint32_t) prev.outputs.size();
+    if (curr_n >= HTP_OP_MAX_OUTPUTS || curr_n + 2 >= HTP_OP_MAX_INPUTS) return false;
+
+    const ggml_tensor * w0     = prev.inputs[0];
+    const ggml_tensor * x      = prev.inputs[curr_n];
+    const ggml_tensor * ids    = prev.inputs[curr_n + 1];
+    const ggml_tensor * w_in   = cur.src0();
+    const ggml_tensor * x_in   = cur.src1();
+    const ggml_tensor * ids_in = cur.node->src[2];
+    const ggml_tensor * d_in   = cur.dst();
+
+    auto * orig_kparams = prev.kparams<htp_mm_kernel_params>();
+    const auto * cur_kparams = cur.kparams<htp_mm_kernel_params>();
+
+    if (x_in != x || ids_in != ids || w_in->type != w0->type ||
+        w_in->ne[0] != w0->ne[0] || w_in->ne[2] != w0->ne[2] ||
+        htp_tensor_is_tiled(w_in) != htp_tensor_is_tiled(w0) ||
+        orig_kparams->n_hmx != cur_kparams->n_hmx) {
+        return false;
+    }
+
+    struct htp_mm_kernel_params kparams;
+    if (d_in->nb[1] == prev.node->nb[1]) {
+        kparams = *orig_kparams;
+        kparams.n_weights = curr_n + 1;
+    } else if (!htp_kparams_solve_mmidnx(sess, w0, x, d_in, curr_n + 1, &kparams)) {
+        return false;
+    }
+    if (!is_supported_mul_mat_id_nx_kernel(w0, &kparams)) return false;
+
+    prev.opcode = HTP_OP_MUL_MAT_ID_NX;
+    prev.inputs.insert(prev.inputs.begin() + curr_n, w_in);
+    prev.outputs.push_back(d_in);
+    prev.fused.push_back(cur.node);
+    prev.set_kparams(kparams);
+    return true;
+}
+
+static bool fuse_gdn_cpy(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    GGML_UNUSED(sess);
+    if (prev.opcode != HTP_OP_GATED_DELTA_NET || prev.outputs.size() != 1) return false;
+
+    const ggml_tensor * gdn_out  = prev.dst();
+    const ggml_tensor * cpy_node = cur.node;
+    const ggml_tensor * cpy_src  = cur.src0();
+    const ggml_tensor * cpy_dst  = cur.dst();
+
+    if (gdn_out->type != GGML_TYPE_F32 || cpy_src->type != GGML_TYPE_F32 || cpy_dst->type != GGML_TYPE_F32) return false;
+    if ((gdn_out->flags & GGML_TENSOR_FLAG_OUTPUT) || (cpy_node->flags & GGML_TENSOR_FLAG_OUTPUT)) return false;
+
+    const ggml_tensor * v = prev.node->src[2];
+    if (!v) return false;
+
+    const int64_t S_v       = v->ne[0];
+    const int64_t H         = v->ne[1];
+    const int64_t n_tokens  = v->ne[2];
+    const int64_t n_seqs    = v->ne[3];
+    const int64_t K         = ggml_get_op_params_i32(prev.node, 0);
+    const size_t  tail_off  = (size_t) S_v * H * n_tokens * n_seqs * sizeof(float);
+    const int64_t D         = S_v * S_v * H;
+    const int64_t n_written = std::min<int64_t>(n_tokens, K);
+
+    if (cpy_src->op != GGML_OP_VIEW || !htp_same_tensor_data(cpy_src->view_src, gdn_out) ||
+        cpy_src->view_offs != tail_off || !ggml_is_contiguous(cpy_src)) {
+        return false;
+    }
+
+    if (cpy_dst->ne[0] != D || cpy_dst->ne[1] != n_seqs || cpy_dst->ne[2] != n_written || !ggml_is_contiguous(cpy_dst)) {
+        return false;
+    }
+
+    prev.outputs.push_back(cpy_dst);
+    prev.fused.push_back(cur.node);
+    return true;
+}
+
+static bool fuse_ssm_conv_silu(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    GGML_UNUSED(sess);
+    if (prev.opcode != HTP_OP_SSM_CONV) return false;
+
+    auto * kparams = prev.kparams<htp_ssm_conv_kernel_params>();
+    if (kparams->silu) return false;
+    if (!htp_match_consumer(prev, cur, /*commutative=*/ false)) return false;
+
+    const ggml_tensor * conv_dst = prev.dst();
+    const ggml_tensor * dst      = cur.dst();
+    if (!ggml_are_same_layout(conv_dst, dst)) return false;
+    if (htp_tensors_overlap(dst, prev.src0()) || htp_tensors_overlap(dst, prev.src1())) return false;
+
+    kparams->silu   = 1;
+    prev.outputs[0] = dst;
+    prev.fused.push_back(cur.node);
+    return true;
+}
+
+static bool fuse_repeat_get_rows(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (prev.opcode != HTP_OP_REPEAT) return false;
+    if (!htp_match_consumer(prev, cur, /*commutative=*/ false)) return false;
+
+    const ggml_tensor * repeat_out  = prev.dst();
+    const ggml_tensor * repeat_src0 = prev.src0();
+    const ggml_tensor * gr_src1     = cur.src1();
+    const ggml_tensor * gr_dst      = cur.dst();
+
+    if (repeat_src0->ne[0] != repeat_out->ne[0] || repeat_src0->ne[1] != repeat_out->ne[1]) return false;
+    if (repeat_src0->type != repeat_out->type) return false;
+
+    struct htp_get_rows_kernel_params kparams;
+    if (!htp_kparams_solve_repeat_get_rows(sess, repeat_src0, gr_src1, gr_dst, &kparams) || kparams.vtcm_size == 0) return false;
+
+    prev.opcode  = HTP_OP_GET_ROWS;
+    prev.inputs  = { repeat_src0, gr_src1 };
+    prev.outputs = { gr_dst };
+    prev.fused.push_back(cur.node);
+    prev.set_kparams(kparams);
+    return true;
+}
+
+static bool fuse_unary_mul_glu(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    htp_op_code glu_op = HTP_OP_INVALID;
+    switch (prev.opcode) {
+        case HTP_OP_UNARY_SILU:       glu_op = HTP_OP_GLU_SWIGLU;      break;
+        case HTP_OP_UNARY_GELU:       glu_op = HTP_OP_GLU_GEGLU;       break;
+        case HTP_OP_UNARY_GELU_QUICK: glu_op = HTP_OP_GLU_GEGLU_QUICK; break;
+        case HTP_OP_UNARY_GELU_ERF:   glu_op = HTP_OP_GLU_GEGLU_ERF;   break;
+        case HTP_OP_UNARY_RELU:       glu_op = HTP_OP_GLU_REGLU;       break;
+        case HTP_OP_UNARY_SIGMOID:    glu_op = HTP_OP_GLU_SIGMOID;     break;
+        default: return false;
+    }
+
+    const ggml_tensor * up = nullptr;
+    if (!htp_match_consumer(prev, cur, /*commutative=*/ true, &up)) return false;
+
+    const ggml_tensor * gate = prev.src0();
+    const ggml_tensor * dst  = cur.dst();
+    if (htp_inputs_overlap_unsafe(prev, dst, up)) return false;
+
+    prev.opcode  = glu_op;
+    prev.inputs  = { gate, up };
+    prev.outputs = { dst };
+    if (!htp_kparams_solve(sess, prev)) return false;
+
+    prev.fused.push_back(cur.node);
+    memset(prev.params, 0, sizeof(prev.params));
+    return true;
+}
+
+static bool fuse_mul_mul_scalar(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (prev.opcode != HTP_OP_MUL || prev.inputs.size() != 2) return false;
+
+    const auto * k0 = prev.kparams<htp_binary_kernel_params>();
+    const auto * k1 = cur.kparams<htp_binary_kernel_params>();
+    if (k0->kernel_type != HTP_BINARY_KERNEL_SCALAR || k1->kernel_type != HTP_BINARY_KERNEL_SCALAR) return false;
+    if (k0->src1_strided || k1->src1_strided) return false;
+
+    const ggml_tensor * src2 = nullptr;
+    if (!htp_match_consumer(prev, cur, /*commutative=*/ false, &src2)) return false;
+
+    const ggml_tensor * src1 = prev.src1();
+    const ggml_tensor * dst  = cur.dst();
+    if (!ggml_are_same_shape(src1, src2) || !htp_same_shape_and_contiguity(prev.dst(), dst)) return false;
+    if (htp_inputs_overlap_unsafe(prev, dst, src2)) return false;
+
+    const size_t vtcm_size = htp_session_vtcm_size(sess);
+    struct htp_binary_kernel_params kparams = *k0;
+    kparams.src2_size       = kparams.src1_size;
+    kparams.rows_per_buffer = 0;
+    if (kparams.src1_size > 65536 || kparams.src1_size + kparams.src2_size > vtcm_size) return false;
+
+    struct htp_binary_vtcm_layout L;
+    htp_binary_vtcm_layout_build(&L, &kparams, vtcm_size);
+    if (L.rows_per_buffer == 0 || L.total_bytes > vtcm_size) return false;
+
+    kparams.rows_per_buffer = L.rows_per_buffer;
+    kparams.vtcm_size       = L.total_bytes;
+
+    prev.inputs.push_back(src2);
+    prev.outputs[0] = dst;
+    prev.fused.push_back(cur.node);
+    prev.set_kparams(kparams);
+    return true;
+}
+
+static bool fuse_binary_nx(const ggml_hexagon_session * sess, htp_opnode & prev, const htp_opnode & cur) {
+    if (prev.opcode != cur.opcode) return false;
+
+    const uint32_t curr_n = (uint32_t) prev.inputs.size();
+    if (curr_n < 2 || curr_n >= HTP_OP_MAX_INPUTS) return false;
+
+    const auto * k0 = prev.kparams<htp_binary_kernel_params>();
+    const auto * k1 = cur.kparams<htp_binary_kernel_params>();
+    const bool k0_ok = (k0->kernel_type == HTP_BINARY_KERNEL_SAME_SHAPE) ||
+                       (k0->kernel_type == HTP_BINARY_KERNEL_CHUNKED && !k0->is_scalar);
+    const bool k1_ok = (k1->kernel_type == HTP_BINARY_KERNEL_SAME_SHAPE) ||
+                       (k1->kernel_type == HTP_BINARY_KERNEL_CHUNKED && !k1->is_scalar);
+    if (!k0_ok || !k1_ok) return false;
+
+    const bool is_comm = (cur.opcode == HTP_OP_ADD || cur.opcode == HTP_OP_MUL);
+    const ggml_tensor * new_src = nullptr;
+    if (!htp_match_consumer(prev, cur, is_comm, &new_src)) return false;
+
+    const ggml_tensor * src0 = prev.inputs[0];
+    const ggml_tensor * dst  = cur.dst();
+    if (!ggml_are_same_shape(src0, prev.inputs[1]) ||
+        !ggml_are_same_shape(src0, new_src) ||
+        !htp_same_shape_and_contiguity(prev.dst(), dst)) {
+        return false;
+    }
+    if (htp_inputs_overlap_unsafe(prev, dst, new_src)) return false;
+
+    const size_t vtcm_size = htp_session_vtcm_size(sess);
+    struct htp_binary_kernel_params kparams = *k0;
+    struct htp_binary_vtcm_layout L;
+    memset(&L, 0, sizeof(L));
+
+    const uint32_t n_srcs = curr_n + 1;
+    if (k0->kernel_type == HTP_BINARY_KERNEL_SAME_SHAPE || k1->kernel_type == HTP_BINARY_KERNEL_SAME_SHAPE) {
+        if (k0->kernel_type != HTP_BINARY_KERNEL_SAME_SHAPE) {
+            kparams = *k1;
+        }
+        kparams.n_srcs          = (uint8_t) n_srcs;
+        kparams.rows_per_buffer = 0;
+
+        auto check_split = [&](const ggml_tensor * t) {
+            if (src0->ne[2] > 1 && t->nb[2] != t->ne[1] * t->nb[1]) kparams.split_at_ne01 = 1;
+            if (src0->ne[3] > 1 && t->nb[3] != t->ne[2] * t->nb[2]) kparams.split_at_ne02 = 1;
+        };
+        for (const auto * in : prev.inputs) check_split(in);
+        check_split(new_src);
+        check_split(dst);
+
+        htp_binary_vtcm_layout_build(&L, &kparams, vtcm_size);
+    }
+
+    if (L.rows_per_buffer == 0 || L.total_bytes == 0 || L.total_bytes > vtcm_size) {
+        bool all_contig = ggml_is_contiguous(new_src) && ggml_is_contiguous(dst);
+        for (const auto * in : prev.inputs) {
+            all_contig = all_contig && ggml_is_contiguous(in);
+        }
+        if (!all_contig) return false;
+
+        const uint32_t total_elems = (uint32_t) ggml_nelements(src0);
+        const uint32_t elem_size   = (uint32_t) ggml_type_size(src0->type);
+        const uint32_t max_threads = htp_session_n_threads(sess);
+        if (!htp_binary_solve_chunked(&L, &kparams, total_elems, elem_size, max_threads, n_srcs, false, vtcm_size)) {
+            return false;
+        }
+    }
+
+    kparams.rows_per_buffer = L.rows_per_buffer;
+    kparams.vtcm_size       = (uint32_t) L.total_bytes;
+
+    prev.inputs.push_back(new_src);
+    prev.outputs[0] = dst;
+    prev.fused.push_back(cur.node);
+    prev.set_kparams(kparams);
+    return true;
+}
+
+bool htp_fusion_can_stack(const struct ggml_tensor * a) {
+    if (!a || !a->src[0] || !a->src[1]) return false;
+    switch (a->op) {
+        case GGML_OP_MUL_MAT:
+            if (!htp_fusion_enabled(HTP_FUSE_MUL_MAT_NX)) return false;
+            if (htp_is_hadamard(a)) return false;
+            if (a->src[0]->ne[2] != 1 || a->src[0]->ne[3] != 1) return false;
+            return ggml_is_quantized(a->src[0]->type);
+        case GGML_OP_MUL_MAT_ID:
+            if (!htp_fusion_enabled(HTP_FUSE_MUL_MAT_ID_NX)) return false;
+            return a->src[2] != nullptr && ggml_is_quantized(a->src[0]->type);
+        default:
+            return false;
+    }
+}
+
+bool htp_fusion_can_stack(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (!htp_fusion_can_stack(b) || a->op != b->op || a->src[1] != b->src[1]) {
+        return false;
+    }
+    if (a->op == GGML_OP_MUL_MAT_ID && a->src[2] != b->src[2]) {
+        return false;
+    }
+    const ggml_tensor * w0 = a->src[0];
+    const ggml_tensor * w1 = b->src[0];
+    return w0->type == w1->type && w0->ne[0] == w1->ne[0] && w0->ne[2] == w1->ne[2];
+}
+
+int htp_fusion_max(const struct ggml_cgraph * graph, int idx) {
+    const int n = graph->n_nodes;
+    const ggml_op op0 = graph->nodes[idx]->op;
+    if (op0 != GGML_OP_ADD && op0 != GGML_OP_NORM && op0 != GGML_OP_RMS_NORM) {
+        return 1;
+    }
+
+    enum ggml_op ops[HTP_OP_MAX_FUSED];
+    ops[0] = op0;
+
+    int f = idx + 1;
+    while (f < n && f < idx + HTP_OP_MAX_FUSED) {
+        const ggml_op op = graph->nodes[f]->op;
+        if (op != GGML_OP_ADD && op != GGML_OP_MUL && op != GGML_OP_SCALE &&
+            op != GGML_OP_NORM && op != GGML_OP_RMS_NORM) {
+            break;
+        }
+        ops[f - idx] = op;
+        f++;
+    }
+
+    f -= idx;
+    for (; f > 1; f--) {
+        if (ggml_can_fuse(graph, idx, ops, f)) {
+            break;
+        }
+    }
+    return f;
+}
+
+void htp_fusion_tag_graph(struct ggml_cgraph * graph) {
+    for (int i = 0; i < graph->n_nodes; i++) {
+        auto * extra = (htp_tensor_extra *) graph->nodes[i]->extra;
+        if (!extra) continue;
+
+        extra->flags &= ~HTP_TENSOR_FUSEABLE;
+
+        const ggml_op op = graph->nodes[i]->op;
+        if ((op == GGML_OP_RMS_NORM || op == GGML_OP_SCALE || op == GGML_OP_REPEAT ||
+             op == GGML_OP_UNARY    || op == GGML_OP_ADD   || op == GGML_OP_SUB    ||
+             op == GGML_OP_MUL      || op == GGML_OP_DIV   || op == GGML_OP_MUL_MAT_ID ||
+             op == GGML_OP_SSM_CONV ||
+             (op == GGML_OP_MUL_MAT && !htp_is_hadamard(graph->nodes[i]))) &&
+            ggml_node_has_n_uses(graph, i, 1)) {
+            extra->flags |= HTP_TENSOR_FUSEABLE;
+        }
+    }
+}
+
+bool htp_fusion_try_fuse(const struct ggml_hexagon_session * sess, struct htp_opnode & prev, const struct htp_opnode & cur) {
+    if (!opt_opfusion) return false;
+
+    switch (cur.opcode) {
+        case HTP_OP_ADD:
+            return (htp_fusion_enabled(HTP_FUSE_ALLREDUCE_ADD) && fuse_allreduce_add(sess, prev, cur)) ||
+                   (htp_fusion_enabled(HTP_FUSE_MUL_MAT_ADD)   && fuse_mul_mat_add(sess, prev, cur))   ||
+                   (htp_fusion_enabled(HTP_FUSE_BINARY_NX)     && fuse_binary_nx(sess, prev, cur));
+        case HTP_OP_MUL:
+            return (htp_fusion_enabled(HTP_FUSE_RMS_NORM_MUL)   && fuse_rms_norm_mul(sess, prev, cur))   ||
+                   (htp_fusion_enabled(HTP_FUSE_UNARY_MUL_GLU)  && fuse_unary_mul_glu(sess, prev, cur))  ||
+                   (htp_fusion_enabled(HTP_FUSE_MUL_MUL_SCALAR) && fuse_mul_mul_scalar(sess, prev, cur)) ||
+                   (htp_fusion_enabled(HTP_FUSE_BINARY_NX)      && fuse_binary_nx(sess, prev, cur));
+        case HTP_OP_SUB:
+        case HTP_OP_DIV:
+            return htp_fusion_enabled(HTP_FUSE_BINARY_NX) && fuse_binary_nx(sess, prev, cur);
+        case HTP_OP_SCALE:
+            return htp_fusion_enabled(HTP_FUSE_RMS_NORM_SCALE) && fuse_rms_norm_scale(sess, prev, cur);
+        case HTP_OP_UNARY_SILU:
+            return htp_fusion_enabled(HTP_FUSE_SSM_CONV_SILU) && fuse_ssm_conv_silu(sess, prev, cur);
+        case HTP_OP_GET_ROWS:
+            return htp_fusion_enabled(HTP_FUSE_REPEAT_GET_ROWS) && fuse_repeat_get_rows(sess, prev, cur);
+        case HTP_OP_CPY:
+            return htp_fusion_enabled(HTP_FUSE_GDN_CPY) && fuse_gdn_cpy(sess, prev, cur);
+        case HTP_OP_MUL_MAT:
+            return htp_fusion_enabled(HTP_FUSE_MUL_MAT_NX) && fuse_mul_mat_nx(sess, prev, cur);
+        case HTP_OP_MUL_MAT_ID:
+            return htp_fusion_enabled(HTP_FUSE_MUL_MAT_ID_NX) && fuse_mul_mat_id_nx(sess, prev, cur);
+        default:
+            return false;
+    }
+}
@@ -0,0 +1,33 @@
+#ifndef HTP_FUSION_H
+#define HTP_FUSION_H
+
+#include "htp-common.h"
+
+enum htp_fusion_flags {
+    HTP_FUSE_ALLREDUCE_ADD   = (1 << 1),  // 2
+    HTP_FUSE_RMS_NORM_MUL    = (1 << 2),  // 4
+    HTP_FUSE_MUL_MAT_ADD     = (1 << 3),  // 8
+    HTP_FUSE_MUL_MAT_NX      = (1 << 4),  // 16
+    HTP_FUSE_MUL_MAT_ID_NX   = (1 << 5),  // 32
+    HTP_FUSE_GDN_CPY         = (1 << 6),  // 64
+    HTP_FUSE_REPEAT_GET_ROWS = (1 << 7),  // 128
+    HTP_FUSE_UNARY_MUL_GLU   = (1 << 8),  // 256
+    HTP_FUSE_MUL_MUL_SCALAR  = (1 << 9),  // 512
+    HTP_FUSE_BINARY_NX       = (1 << 10), // 1024
+    HTP_FUSE_RMS_NORM_SCALE  = (1 << 11), // 2048
+    HTP_FUSE_SSM_CONV_SILU   = (1 << 12), // 4096
+};
+
+static inline bool htp_fusion_enabled(int flag) {
+    if (opt_opfusion <= 0) return false;
+    if (opt_opfusion == 1) return true; // 1 enables all
+    return (opt_opfusion & flag) != 0;
+}
+
+bool htp_fusion_can_stack(const struct ggml_tensor * a);
+bool htp_fusion_can_stack(const struct ggml_tensor * a, const struct ggml_tensor * b);
+int  htp_fusion_max(const struct ggml_cgraph * graph, int idx);
+void htp_fusion_tag_graph(struct ggml_cgraph * graph);
+bool htp_fusion_try_fuse(const struct ggml_hexagon_session * sess, struct htp_opnode & prev, const struct htp_opnode & cur);
+
+#endif // HTP_FUSION_H
--- a/Show More
+++ b/Show More