llama_moe_cache: run missed experts on CPU, add tuning script (#30966 )

* moe-cache : compute part of the misses on the CPU --moe-cache-cpu-share F computes a share F of the MoE cache misses on the CPU from the host experts instead of uploading them. The graph builds the expert FFN twice: on the device from the cache slots, where the misses left to the CPU read zero slots, and on the CPU from the host experts, where the experts computed on the device have id -1. The two results are added before the down bias. - ggml-cpu: MUL_MAT_ID skips negative ids and writes zero rows - sched: the copy callback is called for all the inputs of a split, and before the host sync when there are no events - the slot lookup callback reads the inputs of the CPU part together with the selected experts, so the CPU part does not wait for the device, and its result is copied back asynchronously - each token uses distinct zero slots, mul_mat_id does not support an expert selected twice by a token - LLAMA_MOE_CACHE_TRACE writes the selected experts of each lookup, for a simulator of the cache * scripts : add moe-cache-tune.py to pick the MoE cache arguments The script wraps a llama-server command: it starts the server at two cache sizes and three CPU shares, fits the upload and CPU speeds of the machine, and replays the routing of the runs through a copy of the cache logic to predict other sizes and shares. It prints the recommended --moe-cache-mib and --moe-cache-cpu-share. docs/moe-cache.md describes the MoE cache arguments and the script. * add am17an as codeowner for MoE cache * moe-cache : upload all the misses until the cache is full With a high --moe-cache-cpu-share, few misses are uploaded, so an empty cache fills slowly and more misses than needed go to the CPU. The CPU share now starts once all the slots of a cache group hold an expert. moe-cache-tune.py replays the same rule. * add backend sync for async transfer queue * moe-cache : let the scheduler copy the results of the CPU part * Fix MSVC compilation * remove sync change * llama-graph : extract the CPU share condition into can_cpu_share * llama-graph : make llm_graph_params::moe_cache const The graph construction no longer updates the MoE cache directly: the CPU parts of the layers are accumulated in the graph result and applied to the cache in llama_context::process_ubatch, after the graph inputs are set and before the graph is computed. Assisted-by: dsh:llama.cpp/ggml-org/GLM-5.3-Flash-GGUF (Mac Studio, M5 Ultra, 256GB) * scripts : replay the ClassLRU policy in moe-cache-tune.py * moe-cache : pick CPU misses by CRF instead of random (#31853) * moe-cache : remove the backend logic of the CPU part from the graph The CPU part reads the host experts, so the scheduler places it on the CPU without set_tensor_backend. The cache gives no CPU map for batches that a device would offload. Graphs with a CPU part look up their own slot map, so the graph does not register the part with the cache. ggml-backend : the inputs of the next split that are already computed are copied to host memory before the current split runs, so a CPU split does not wait for the device split before it. This replaces the stash of the cache, and the copy callback is again called for host weights only. * llama-graph : return the MoE cache slots and CPU ids in a struct The scheduler prefetch no longer checks that the next split runs on another backend: the inputs of a split are never on its own backend, so the input loop already skips that case. --------- Co-authored-by: Gaurav Garg <gaugarg@nvidia.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Alexey Kopytko <alexey@kopytko.com>
sycl: add f16 support for SWIGLU_OAI (#30857 )
2026-11-07 11:27:38 -06:00 · 2026-11-08 00:20:58 +08:00 · 2026-11-07 09:02:58 -05:00 · 2026-11-07 13:04:13 +02:00 · 2026-11-07 13:03:38 +02:00 · 2026-11-07 10:01:27 +02:00
653 changed files with 75595 additions and 26875 deletions
@@ -84,6 +84,9 @@ RUN apt-get update \
    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
    && find /var/cache -type f -delete

+# Some hosts (e.g. GKE) only mount the driver here, and CUDA 12.8.1-12.9.1 base images don't set these paths
+ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/nvidia/lib:${LD_LIBRARY_PATH}
+
 COPY --from=build /app/lib/ /app

 ### Full
@@ -64,7 +64,7 @@ RUN mkdir -p /app/full \
    && cp requirements.txt /app/full \
    && cp .devops/tools.sh /app/full/tools.sh

-FROM docker.io/intel/oneapi-toolkit:$ONEAPI_VERSION AS base
+FROM docker.io/intel/deep-learning-essentials:$ONEAPI_VERSION AS base

 ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
@@ -79,6 +79,8 @@ LABEL org.opencontainers.image.created=$BUILD_DATE \
      org.opencontainers.image.url=$IMAGE_URL \
      org.opencontainers.image.source=$IMAGE_SOURCE

+COPY --from=build /opt/intel/oneapi/dnnl /opt/intel/oneapi/dnnl
+
 #Following versions are for multiple GPUs, since 26.x has known issue:
 #   https://github.com/ggml-org/llama.cpp/issues/21747,
 #   https://github.com/intel/compute-runtime/issues/921.
@@ -100,56 +100,56 @@ runs:
          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
          echo "CUDA_PATH_V13_1=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8

-    - name: Install Cuda Toolkit 13.4 for x64
-      if: ${{ inputs.cuda_version == '13.4' && inputs.cuda_arch == 'x64' }}
+    - name: Install Cuda Toolkit 13.5 for x64
+      if: ${{ inputs.cuda_version == '13.5' && inputs.cuda_arch == 'x64' }}
      shell: pwsh
      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5"
          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.4.59-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-13.4.49-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.4.59-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-13.4.59-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-13.7.0.27-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.4.59-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-13.4.49-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-13.4.49-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-13.4.49-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cccl/windows-x86_64/cccl-windows-x86_64-13.3.4.2.1-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_crt-windows-x86_64-13.4.59-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_cudart-windows-x86_64-13.4.49-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvcc-windows-x86_64-13.4.59-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvrtc-windows-x86_64-13.4.59-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libcublas-windows-x86_64-13.7.0.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libnvvm-windows-x86_64-13.4.59-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvtx-windows-x86_64-13.4.49-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_profiler_api-windows-x86_64-13.4.49-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\visual_studio_integration-windows-x86_64-13.4.49-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cccl-windows-x86_64-13.3.4.2.1-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V13_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-13.10.0.20-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-13.5.25-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-13.5.25-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-13.5.25-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cccl/windows-x86_64/cccl-windows-x86_64-13.3.5.0.6-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_crt-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_cudart-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_nvcc-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_nvrtc-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\libcublas-windows-x86_64-13.10.0.20-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\libnvvm-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_nvtx-windows-x86_64-13.5.25-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_profiler_api-windows-x86_64-13.5.25-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\visual_studio_integration-windows-x86_64-13.5.25-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cccl-windows-x86_64-13.3.5.0.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V13_5=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8

-    - name: Install Cuda Toolkit 13.4 for ARM64
-      if: ${{ inputs.cuda_version == '13.4' && inputs.cuda_arch == 'arm64' }}
+    - name: Install Cuda Toolkit 13.5 for ARM64
+      if: ${{ inputs.cuda_version == '13.5' && inputs.cuda_arch == 'arm64' }}
      shell: pwsh
      run: |
-          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
+          mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5"
          choco install unzip -y
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cccl/windows-x86_64/cccl-windows-x86_64-13.3.4.2.1-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.4.59-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.4.59-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.4.59-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-arm64/cuda_cudart-windows-arm64-13.4.49-archive.zip"
-          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-arm64/libcublas-windows-arm64-13.7.0.27-archive.zip"
-          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4"
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cccl-windows-x86_64-13.3.4.2.1-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_crt-windows-x86_64-13.4.59-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_nvcc-windows-x86_64-13.4.59-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libnvvm-windows-x86_64-13.4.59-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\cuda_cudart-windows-arm64-13.4.49-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\libcublas-windows-arm64-13.7.0.27-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" /E /I /H /Y
-          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
-          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
-          echo "CUDA_PATH_V13_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cccl/windows-x86_64/cccl-windows-x86_64-13.3.5.0.6-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_crt/windows-x86_64/cuda_crt-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libnvvm/windows-x86_64/libnvvm-windows-x86_64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-arm64/cuda_cudart-windows-arm64-13.5.37-archive.zip"
+          curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-arm64/libcublas-windows-arm64-13.10.0.20-archive.zip"
+          unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5"
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cccl-windows-x86_64-13.3.5.0.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_crt-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_nvcc-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\libnvvm-windows-x86_64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\cuda_cudart-windows-arm64-13.5.37-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\libcublas-windows-arm64-13.10.0.20-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" /E /I /H /Y
+          echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
+          echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
+          echo "CUDA_PATH_V13_5=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.5" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8
@@ -13,6 +13,8 @@ on:
      '**/*.hpp',
      '**/*.c',
      '**/*.cpp',
+      '**/*.py',
+      '**/requirements*.txt',
    ]

  pull_request:
@@ -24,7 +26,9 @@ on:
      '**/*.h',
      '**/*.hpp',
      '**/*.c',
-      '**/*.cpp'
+      '**/*.cpp',
+      '**/*.py',
+      '**/requirements*.txt',
    ]

 cache-mode: none
@@ -69,7 +73,7 @@ jobs:
        run: |
          python3 -m venv .venv
          source .venv/bin/activate
-          pip install jinja2==3.1.6
+          pip install -r requirements/requirements-convert_hf_to_gguf.txt

      - name: ccache-buckets-restore
        uses: ./.github/actions/ccache-buckets
@@ -134,7 +138,7 @@ jobs:
        run: |
          python3 -m venv .venv
          source .venv/Scripts/activate
-          pip install jinja2==3.1.6
+          pip install -r requirements/requirements-convert_hf_to_gguf.txt

      - name: Build
        id: cmake_build
@@ -156,3 +160,13 @@ jobs:
          source .venv/Scripts/activate
          cd build
          ctest -C Release -L behavior -LE 'model|backend' --verbose --timeout 900
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: behavior-check-windows-2025
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
@@ -0,0 +1,160 @@
+name: CI (apple - daily)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  schedule:
+    - cron: '0 0 * * *'
+
+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
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  macos-latest-ios-xcode:
+    runs-on: macos-latest
+
+    permissions:
+      actions: write
+      contents: read
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Setup Xcode
+        uses: ggml-org/setup-xcode@v1
+        with:
+          xcode-version: latest-stable
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          ./build-xcframework.sh
+
+      - name: Upload xcframework artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+          retention-days: 1
+
+      - name: Build Xcode project
+        run: |
+          xcodebuild -downloadPlatform iOS
+          xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
+
+  macos-latest-tvos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=tvOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+  macos-latest-visionos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=visionOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+  macos-latest-swift:
+    runs-on: macos-latest
+    needs: macos-latest-ios-xcode
+
+    permissions:
+      actions: read
+      contents: read
+
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Download xcframework artifact
+        uses: actions/download-artifact@v7
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+
+      - name: Build llama.cpp with CMake
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
@@ -107,140 +107,3 @@ jobs:
          # ref: https://github.com/ggml-org/llama.cpp/pull/19802#issuecomment-4013704023
          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
-
-    permissions:
-      actions: write
-      contents: read
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v6
-
-      - name: Setup Xcode
-        uses: ggml-org/setup-xcode@v1
-        with:
-          xcode-version: latest-stable
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=iOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          ./build-xcframework.sh
-
-      - name: Upload xcframework artifact
-        uses: actions/upload-artifact@v6
-        with:
-          name: llama-xcframework
-          path: build-apple/llama.xcframework/
-          retention-days: 1
-
-      - name: Build Xcode project
-        run: |
-          xcodebuild -downloadPlatform iOS
-          xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
-
-  macos-latest-tvos:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_COMMON=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=tvOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
-  macos-latest-visionos:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_COMMON=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=visionOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
-  macos-latest-swift:
-    runs-on: macos-latest
-    needs: macos-latest-ios-xcode
-
-    permissions:
-      actions: read
-      contents: read
-
-    strategy:
-      matrix:
-        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Download xcframework artifact
-        uses: actions/download-artifact@v7
-        with:
-          name: llama-xcframework
-          path: build-apple/llama.xcframework/
-
-      - name: Build llama.cpp with CMake
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
@@ -1,57 +0,0 @@
-name: Build relocatable cmake package
-on:
-  workflow_dispatch:
-  workflow_call:
-
-cache-mode: none
-permissions:
-  contents: read
-
-jobs:
-  linux:
-    runs-on: [self-hosted, Linux, fast]
-    steps:
-      - uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: Build
-        run: |
-          PREFIX="$(pwd)"/inst
-          cmake -S . -B build \
-                -DCMAKE_PREFIX_PATH="$PREFIX" \
-                -DLLAMA_OPENSSL=OFF \
-                -DLLAMA_BUILD_TESTS=OFF \
-                -DLLAMA_BUILD_TOOLS=OFF \
-                -DLLAMA_BUILD_EXAMPLES=OFF \
-                -DLLAMA_BUILD_APP=OFF \
-                -DLLAMA_BUILD_IS_DEV=OFF \
-                -DCMAKE_BUILD_TYPE=Release
-          cmake --build build --config Release -j $(nproc)
-          cmake --install build --prefix "$PREFIX" --config Release
-
-          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          build_commit=$(git rev-parse --short HEAD | xargs)
-          build_number=$(git rev-list --count HEAD | xargs)
-
-          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
-          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
-          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
-          build_version="$major.$minor.$patch"
-
-          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
-                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
-                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")
-
-          for check in "${checks[@]}"; do
-              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
-                  echo "Checking llama-config.cmake version... \"$check\" failed!"
-                  exit 1
-              fi
-          done
-
-          echo "Checking llama-config.cmake version... success."
-
-          cd examples/simple-cmake-pkg
-          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
-          cmake --build build -j $(nproc)
@@ -7,7 +7,6 @@ on:
      - master
    paths: [
      '.github/workflows/build-cpu.yml',
-      '.github/workflows/build-cmake-pkg.yml',
      '**/CMakeLists.txt',
      '**/.cmake',
      '**/*.h',
@@ -20,7 +19,6 @@ on:
    types: [opened, synchronize, reopened]
    paths: [
      '.github/workflows/build-cpu.yml',
-      '.github/workflows/build-cmake-pkg.yml',
      'ggml/src/ggml-rpc/**',
      '**/CMakeLists.txt',
      '**/.cmake',
@@ -47,9 +45,6 @@ env:
  LLAMA_ARG_LOG_TIMESTAMPS: 1

 jobs:
-  build-cmake-pkg:
-    uses: ./.github/workflows/build-cmake-pkg.yml
-
  ubuntu:
    strategy:
      matrix:
@@ -242,3 +237,79 @@ jobs:
          older: 5m
          min: 1
          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
+  build-cmake-pkg:
+    runs-on: ubuntu-24.04
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - 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: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: build-cmake-pkg
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
+      - name: Build
+        run: |
+          PREFIX="$(pwd)"/inst
+          cmake -S . -B build \
+                -DCMAKE_PREFIX_PATH="$PREFIX" \
+                -DLLAMA_OPENSSL=OFF \
+                -DLLAMA_BUILD_TESTS=OFF \
+                -DLLAMA_BUILD_TOOLS=OFF \
+                -DLLAMA_BUILD_EXAMPLES=OFF \
+                -DLLAMA_BUILD_APP=OFF \
+                -DLLAMA_BUILD_IS_DEV=OFF \
+                -DCMAKE_BUILD_TYPE=Release
+          cmake --build build --config Release -j $(nproc)
+          cmake --install build --prefix "$PREFIX" --config Release
+
+          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
+          build_commit=$(git rev-parse --short HEAD | xargs)
+          build_number=$(git rev-list --count HEAD | xargs)
+
+          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          build_version="$major.$minor.$patch"
+
+          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
+                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
+                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")
+
+          for check in "${checks[@]}"; do
+              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
+                  echo "Checking llama-config.cmake version... \"$check\" failed!"
+                  exit 1
+              fi
+          done
+
+          echo "Checking llama-config.cmake version... success."
+
+          cd examples/simple-cmake-pkg
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
+          cmake --build build -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: build-cmake-pkg
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
@@ -40,12 +40,12 @@ jobs:
          - cuda: '12.4'
            arch: x64
            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
-          - cuda: '13.4'
+          - cuda: '13.5'
            arch: x64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
-          - cuda: '13.4'
+            defines: ''
+          - cuda: '13.5'
            arch: arm64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3 -DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -43,13 +43,12 @@ jobs:
    strategy:
      matrix:
        include:
-        # thread and address doesn't run properly on some self hosted machines, so run it on Github instead
        - sanitizer: ADDRESS
-          machine: ubuntu-24.04
+          machine: ubuntu-26.04
        - sanitizer: THREAD
-          machine: ubuntu-24.04
+          machine: ubuntu-26.04
        - sanitizer: UNDEFINED
-          machine: [self-hosted, X64, Linux, CPU]
+          machine: ubuntu-26.04

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

@@ -58,31 +57,24 @@ jobs:
        id: checkout
        uses: actions/checkout@v6

-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.24
-      #   if: ${{ matrix.sanitizer != 'UNDEFINED' }}
-      #   with:
-      #     key: ctest-${{ matrix.sanitizer }}-ubuntu-24.04
-      #     variant: ccache
-      #     evict-old-files: 1d
-      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.24
+        with:
+          restore: false
+          save: false

-      # with UNDEFINED sanitizer, we have to build in Debug to avoid GCC 13 false-positive warnings
-      - name: Build (undefined)
-        id: cmake_build_undefined
-        if: ${{ matrix.sanitizer == 'UNDEFINED' }}
-        run: |
-          cmake -B build \
-            -DCMAKE_BUILD_TYPE=Debug \
-            -DLLAMA_FATAL_WARNINGS=ON \
-            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON
-
-          cmake --build build --config Debug -j $(nproc)
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
+        with:
+          key: ctest-sanitizer-${{ matrix.sanitizer }}
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache

      - name: Build
        id: cmake_build
-        if: ${{ matrix.sanitizer == 'ADDRESS' }}
+        if: ${{ matrix.sanitizer == 'ADDRESS' || matrix.sanitizer == 'UNDEFINED' }}
        run: |
          cmake -B build \
            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
@@ -103,6 +95,18 @@ jobs:

          cmake --build build --config RelWithDebInfo -j $(nproc)

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: ctest-sanitizer-${{ matrix.sanitizer }}
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Test
        id: cmake_test
        run: |
@@ -46,18 +46,55 @@ env:

 jobs:
  cpu-x64-high-perf:
-    runs-on: [self-hosted, Linux, X64, CPU]
+    runs-on: hf-jobs-cpu-upgrade:ubuntu26_04
+
+    # note: this runner has 8 CPUs, but std::thread::hardware_concurrency reports the cores of the
+    #   host (https://github.com/ggml-org/llama.cpp/pull/31439#issuecomment-6683386059)
+    env:
+      TEST_ARG_HARDWARE_CONCURRENCY: 8
+      TEST_ARG_PARALLEL_WORKERS: 4

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v6

+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt update
+          sudo apt install -y build-essential cmake libssl-dev time unzip wget python3 python3-venv python3-dev
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.24
+        with:
+          restore: false
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        with:
+          key: self-hosted-cpu-x64-high-perf
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
      - name: Test
        id: ggml-ci
        run: |
          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp/$RUNNER_NAME ~/mnt/llama.cpp

+      - 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: self-hosted-cpu-x64-high-perf
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
  cpu-arm64-high-perf-graviton4:
    runs-on: ah-ubuntu_24_04-c8g_8x

@@ -92,8 +92,8 @@ jobs:
            { "tag": "cpu", "dockerfile": ".devops/s390x.Dockerfile", "platforms": "linux/s390x", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-s390x", "prebuilt_ui": true },
            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "12.8.1", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
-            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.4.1", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
-            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.4.1", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.5.0", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda.Dockerfile", "cuda_version": "13.5.0", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
            { "tag": "musa", "dockerfile": ".devops/musa.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
            { "tag": "intel", "dockerfile": ".devops/intel.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
            { "tag": "vulkan", "dockerfile": ".devops/vulkan.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },
@@ -11,7 +11,8 @@ on:
      'tests/fusion/**',
      'tests/test-fusion.cpp',
      'tests/test-llama-archs.cpp',
-      'src/llama-graph.cpp',
+      'src/llama-graph*',
+      'src/llama-kv-cache*',
      'src/llama-model*',
      'src/models/**'
    ]
@@ -24,7 +25,8 @@ on:
      'tests/fusion/**',
      'tests/test-fusion.cpp',
      'tests/test-llama-archs.cpp',
-      'src/llama-graph.cpp',
+      'src/llama-graph*',
+      'src/llama-kv-cache*',
      'src/llama-model*',
      'src/models/**'
    ]
@@ -145,6 +147,11 @@ jobs:
          rm -rf build-ci-models && mkdir -p build-ci-models
          ./build/bin/test-llama-archs -o build-ci-models

+      - name: Test KV shift
+        id: test_kv_shift
+        run: |
+          ./build/bin/test-llama-archs --kv-shift
+
      - name: Test fusion
        id: test_fusion
        run: |
@@ -293,6 +300,11 @@ jobs:
      #   run: |
      #     ./build/bin/test-fusion --models build-ci-models --device Vulkan0 --check tests/fusion/Vulkan.csv

+      - name: Test KV shift
+        id: test_kv_shift
+        run: |
+          ./build/bin/test-llama-archs --kv-shift
+
      - name: Test archs
        id: test_archs
        run: |
@@ -0,0 +1,37 @@
+name: "PR video demo label"
+
+on:
+  pull_request_target:
+    types: [opened, edited, reopened, synchronize]
+
+cache-mode: none
+permissions:
+  pull-requests: write
+  issues: write
+
+jobs:
+  video-demo-label:
+    runs-on: ubuntu-slim
+    steps:
+      - name: Add or remove the video demo label
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          PR_URL: ${{ github.event.pull_request.html_url }}
+          PR_NUMBER: ${{ github.event.pull_request.number }}
+          # passed through env, so the body is only matched, never run, and is safe to read from a fork
+          BODY: ${{ github.event.pull_request.body }}
+          HAS_LABEL: ${{ contains(github.event.pull_request.labels.*.name, 'video demo') }}
+        run: |
+          # sections written by demo:pr (see tools/ui/demo/cli.ts): the managed comment
+          # markers, or a preview linking to a demo video uploaded for this PR
+          has_demo=false
+          if { grep -qF '<!-- demo:start -->' <<< "$BODY" && grep -qF '<!-- demo:end -->' <<< "$BODY"; } ||
+            grep -qE "/resolve/videos/pr/$PR_NUMBER/[0-9a-f]{10}(-[0-9]+)?\.mp4" <<< "$BODY"; then
+            has_demo=true
+          fi
+
+          if [ "$has_demo" = true ] && [ "$HAS_LABEL" = false ]; then
+            gh pr edit "$PR_URL" --add-label 'video demo'
+          elif [ "$has_demo" = false ] && [ "$HAS_LABEL" = true ]; then
+            gh pr edit "$PR_URL" --remove-label 'video demo'
+          fi
@@ -177,8 +177,8 @@ jobs:
            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
            - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
            - [Ubuntu x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-12.8-x64.tar.gz) - [CUDA 12.8 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-12.8-x64.tar.gz)
-            - [Ubuntu x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-x64.tar.gz) - [CUDA 13.4 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-x64.tar.gz)
-            - [Ubuntu arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-arm64.tar.gz) - [CUDA 13.4 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.4-arm64.tar.gz)
+            - [Ubuntu x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.5-x64.tar.gz) - [CUDA 13.5 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.5-x64.tar.gz)
+            - [Ubuntu arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.5-arm64.tar.gz) - [CUDA 13.5 libraries](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-${{ steps.tag.outputs.name }}-bin-ubuntu-cuda-13.5-arm64.tar.gz)
            - [Ubuntu x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-10.0-x64.tar.gz)
            - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-2026.4.1-x64.tar.gz)
            - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
@@ -194,8 +194,8 @@ jobs:
            - [Windows arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-arm64.zip)
            - [Windows arm64 (OpenCL Adreno)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-opencl-adreno-arm64.zip)
            - [Windows x64 (CUDA 12)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-12.4-x64.zip) - [CUDA 12.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-12.4-x64.zip)
-            - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.4-x64.zip) - [CUDA 13.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.4-x64.zip)
-            - [Windows arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.4-arm64.zip) - [CUDA 13.4 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.4-arm64.zip)
+            - [Windows x64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.5-x64.zip) - [CUDA 13.5 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.5-x64.zip)
+            - [Windows arm64 (CUDA 13)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cuda-13.5-arm64.zip) - [CUDA 13.5 DLLs](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/cudart-llama-bin-win-cuda-13.5-arm64.zip)
            - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
            - [Windows arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-arm64.zip)
            - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-2026.4.1-x64.zip)
@@ -345,14 +345,14 @@ jobs:
            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
          - build: 'x64'
            os: ubuntu-24.04
-            cuda: '13.4.1'
-            label: '13.4'
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            cuda: '13.5.0'
+            label: '13.5'
+            defines: ''
          - build: 'arm64'
            os: ubuntu-24.04-arm
-            cuda: '13.4.1'
-            label: '13.4'
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
+            cuda: '13.5.0'
+            label: '13.5'
+            defines: ''

    runs-on: ${{ matrix.os }}
    container: nvidia/cuda:${{ matrix.cuda }}-devel-ubuntu24.04
@@ -1211,12 +1211,12 @@ jobs:
          - cuda: '12.4'
            arch: x64
            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
-          - cuda: '13.4'
+          - cuda: '13.5'
            arch: x64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3'
-          - cuda: '13.4'
+            defines: ''
+          - cuda: '13.5'
            arch: arm64
-            defines: '-DGGML_CUDA_CCCL_VERSION=v3.4.3 -DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -1,4 +1,4 @@
-# date: Wed Oct 28 12:38:26 EET 2026
+# date: Wed Nov  4 12:03:31 EET 2026
 # this file is auto-generated by scripts/gen-authors.sh

 Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>
@@ -48,6 +48,7 @@ Abheek Gulati <abheekg@hotmail.com>
 abhijain1204fujitsu <139222713+abhijain1204fujitsu@users.noreply.github.com>
 Abhijit Ramesh <abhijitramesh2k@gmail.com>
 abhijitb11 <113058133+abhijitb11@users.noreply.github.com>
+Abhijith Jajam <129675815+jajamabhijith@users.noreply.github.com>
 Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
 Abhinay Krishna <abhinaykrishna60@gmail.com>
 Abhiram <78226909+geckguy@users.noreply.github.com>
@@ -203,14 +204,17 @@ Ankit Khandelwal <ankk98@gmail.com>
 Ankur Verma <31362771+ankurvdev@users.noreply.github.com>
 anon998 <131767832+anon998@users.noreply.github.com>
 Anri Lombard <anri.m.lombard@gmail.com>
+Anthony Maltsev <77708730+anthonymaltsev@users.noreply.github.com>
 Anthony Umfer <aumfer@gmail.com>
 Anthony Van de Gejuchte <anthonyvdgent@gmail.com>
 Anthony Wang <anthonywang903@gmail.com>
 antichristHater <142441588+antichristHater@users.noreply.github.com>
 Antoine Viallon <antoine@lesviallon.fr>
+Anton I <inflyer@hotmail.com>
 Anton Mitkov <anton_b_mitkov@abv.bg>
 Anton Mitkov <anton.mitkov@codeplay.com>
 Antonis Makropoulos <benuix@gmail.com>
+anubhavfujitsu <Anubhav.Srivastava@fujitsu.com>
 Anudit Nagar <nagaranudit@gmail.com>
 Anuj Attri <anujattri01@gmail.com>
 anujj <ajalota@nvidia.com>
@@ -293,6 +297,7 @@ Bernat Vadell <hounter.caza@gmail.com>
 Bernhard M. Wiedemann <githubbmwprimary@lsmod.de>
 Bert Wagner <github@bertwagner.com>
 Bertay Eren <39909689+bertaye@users.noreply.github.com>
+Bhargav Krish <bkrishnamurt@nvidia.com>
 Bhavik Sharda <10757940+BLSharda@users.noreply.github.com>
 bhubbb <79117352+bhubbb@users.noreply.github.com>
 Bill Sideris <bill88t@feline.gr>
@@ -533,6 +538,7 @@ Don Mahurin <dmahurin@users.noreply.github.com>
 Dong Won Kim <63934649+ddwkim@users.noreply.github.com>
 Donghyeon Jeong <54725479+djeong20@users.noreply.github.com>
 Dongliang Wei <121270393+wdl339@users.noreply.github.com>
+Dongsheng Luo <luods1995@outlook.com>
 Doomsdayrs <38189170+Doomsdayrs@users.noreply.github.com>
 DooWoong Lee (David) <manics99@naver.com>
 DorianRudolph <dorianrudo97@googlemail.com>
@@ -612,6 +618,7 @@ Faez Shakil <faez.shakil@gmail.com>
 fairydreaming <166155368+fairydreaming@users.noreply.github.com>
 Faisal Zaghloul <faisal.zaghloul@gmail.com>
 Faisal Zaghloul <quic_fzaghlou@quicinc.com>
+FamousMonsterr <famousmonster@ya.ru>
 Fan Shupei <dymarkfan@outlook.com>
 FantasyGmm <16450052+FantasyGmm@users.noreply.github.com>
 fanyang <fanyang89@outlook.com>
@@ -654,6 +661,7 @@ Frederik Vogel <Schaltfehler@users.noreply.github.com>
 Fredrik Hultin <noname@nurd.se>
 fredzillman <fzillman@gmail.com>
 frob <rick+github@frob.com.au>
+frogscooper <frogscooper@gmail.com>
 Frost-54 <91168148+Frost-54@users.noreply.github.com>
 Frosty40 <newjordan@gmail.com>
 Funtowicz Morgan <mfuntowicz@users.noreply.github.com>
@@ -694,6 +702,7 @@ Gilad S. <7817232+giladgd@users.noreply.github.com>
 Giovanni Rivera <rivera.giovanni271@gmail.com>
 github-actions[bot] <github-actions[bot]@users.noreply.github.com>
 GittyBurstein <g0534163997@gmail.com>
+Giulianno <lernexbusiness@gmail.com>
 Giuseppe Scrivano <giuseppe@scrivano.org>
 Giuseppe Scrivano <gscrivan@redhat.com>
 GiviMAD <GiviMAD@users.noreply.github.com>
@@ -762,6 +771,7 @@ Henry147147 <44851451+Henry147147@users.noreply.github.com>
 Herman Semenoff <GermanAizek@yandex.ru>
 Herman Semenov <GermanAizek@yandex.ru>
 Hesen Peng <hesen.peng@gmail.com>
+hesiru <siruhe666@gmail.com>
 hey-gm <heitorgm@outlook.com>
 HighDoping <highdoping@gmail.com>
 HimariO <dsfhe49854@gmail.com>
@@ -919,6 +929,7 @@ Jesse Ikonen <jesse.ikonen@gmail.com>
 Jesse Jojo Johnson <williamsaintgeorge@gmail.com>
 Jesse LaRose <jesse@taey.ai>
 Jesse Posner <jesse.posner@gmail.com>
+Jesse Symons <155509533+Torchit1@users.noreply.github.com>
 Jesus Gulfo <jesdga95@gmail.com>
 Jesus Talavera <145992175+jesus-talavera-ibm@users.noreply.github.com>
 Jetson Tan <tanzongyouyi@outlook.com>
@@ -1071,6 +1082,7 @@ Kevin Liu <4396kevinliu@gmail.com>
 Kevin Lo <kevlo@kevlo.org>
 Kevin Pouget <kpouget@redhat.com>
 Kevin Wang <kevmo314@gmail.com>
+Keyu Zhu <inuebisu@outlook.com>
 Khashayar Ghafouri <43180261+khashayarghafouri@users.noreply.github.com>
 khimaros <me@khimaros.com>
 Kilian Hu <90606809+kilian-hu@users.noreply.github.com>
@@ -1078,6 +1090,7 @@ Kilian Krampf <kilian@krampf.de>
 kiltyj <kiltyj@gmail.com>
 Kim S. <polydecay@users.noreply.github.com>
 kimminsu <80271594+kimminsu38oo@users.noreply.github.com>
+Kirill A. Korinsky <kirill@korins.ky>
 KITAITI Makoto <KitaitiMakoto@gmail.com>
 kiwi <122582483+kiwi142857@users.noreply.github.com>
 klosax <131523366+klosax@users.noreply.github.com>
@@ -1136,6 +1149,7 @@ leejet <leejet714@gmail.com>
 Leng Yue <lengyue@lengyue.me>
 Lennart Austenfeld <53152202+l-austenfeld@users.noreply.github.com>
 leo-pony <nengjunma@outlook.com>
+Leon <118652405+LeON-Nie-code@users.noreply.github.com>
 Leon Knauer <git@leonknauer.com>
 Leonard Mosescu <tlemo@users.noreply.github.com>
 leonardHONG <2695316095@qq.com>
@@ -1146,8 +1160,10 @@ lexasub <lexakopp2212@gmail.com>
 lgai-exaone <exaonemodels@lgresearch.ai>
 lhez <lih@qti.qualcomm.com>
 lhez <quic_lih@quicinc.com>
+Li Chuyang <1418587903@qq.com>
 Li Pengzhan <151381994+Lpzhan931@users.noreply.github.com>
 Li Tan <tanliboy@gmail.com>
+li-lizhe <82501526+li-lizhe@users.noreply.github.com>
 liangliangchang <lichang@amd.com>
 liminfei-amd <91481003+liminfei-amd@users.noreply.github.com>
 limitedAtonement <limitedAtonement@users.noreply.github.com>
@@ -1166,6 +1182,7 @@ LoganDark <github@logandark.mozmail.com>
 Loïc Carrère <loic.carrere@gmail.com>
 lon <114724657+longregen@users.noreply.github.com>
 loonerin <132926317+loonerin@users.noreply.github.com>
+Lorenzo Bona <lorenzobona@tuta.com>
 losewayy <1134285612@qq.com>
 LostRuins <39025047+LostRuins@users.noreply.github.com>
 LostRuins Concedo <39025047+LostRuins@users.noreply.github.com>
@@ -1253,6 +1270,7 @@ Masaya, Kato <62578291+msy-kato@users.noreply.github.com>
 mashdragon <122402293+mashdragon@users.noreply.github.com>
 Mason Milburn <masonmilby@gmail.com>
 MasterYi1024 <39848311+MasterYi1024@users.noreply.github.com>
+Maternion <98753158+maternion@users.noreply.github.com>
 Mateusz Charytoniuk <mateusz.charytoniuk@protonmail.com>
 Matheus C. França <matheus-catarino@hotmail.com>
 Matheus Gabriel Alves Silva <matheusgasource@gmail.com>
@@ -1350,6 +1368,7 @@ Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com>
 miyan <1138989048@qq.com>
 mj-shifu <77107165+mj-shifu@users.noreply.github.com>
 mkoker <132301062+mkoker@users.noreply.github.com>
+mmalpe-nv <mmalpe@nvidia.com>
 mmyjona <jonathan.gonse@gmail.com>
 mnehete32 <33429707+mnehete32@users.noreply.github.com>
 Mohamed Elashri <git@elashri.com>
@@ -1368,6 +1387,7 @@ mtmcp <141645996+mtmcp@users.noreply.github.com>
 muggle-stack <promuggle@qq.com>
 Muhammad Saad <saadydev@gmail.com>
 Muhammad Salem <salem.ebo@gmail.com>
+MuoDoo <36472129+MuoDoo@users.noreply.github.com>
 Murilo Santana <mvrilo@gmail.com>
 Musab Gultekin <musabgultekin@users.noreply.github.com>
 musoles <135031143+musoles@users.noreply.github.com>
@@ -1416,10 +1436,12 @@ Nicolas Budyn <16596345+nbud@users.noreply.github.com>
 Nicolas Mowen <nickmowen213@gmail.com>
 Nicolás Pérez <nicolas_perez@brown.edu>
 Nicolò Scipione <nicolo.scipione@codeplay.com>
+Nidhi <nidhisharma16635@gmail.com>
 Nigel Bosch <pnigelb@gmail.com>
 Nigel Jones <nigel.l.jones+git@gmail.com>
 Nik Bogatyrev <50734723+notforu777@users.noreply.github.com>
 Nikhil Jain <nikhil.jain0987@gmail.com>
+Nikhil Sharma <150424968+nikhilfujitsu@users.noreply.github.com>
 Nikita Gordeev <iamfobey@gmail.com>
 Nikita Sarychev <42014488+sARY77@users.noreply.github.com>
 Niklas Korz <niklas@niklaskorz.de>
@@ -1592,6 +1614,7 @@ Raman Shinde <raman.shinde15@gmail.com>
 Rand Xie <randxiexyy29@gmail.com>
 Randall Fitzgerald <randall@dasaku.net>
 Random Fly <renfei8@live.cn>
+randomized <286843573@qq.com>
 rankaiyx <rankaiyx@foxmail.com>
 rankaiyx <rankaiyx@rankaiyx.com>
 RapidMark <32768622+RapidMark@users.noreply.github.com>
@@ -2009,6 +2032,7 @@ Tungsten842 <quantmint@protonmail.com>
 Tushar <ditsuke@protonmail.com>
 tv1wnd <55383215+tv1wnd@users.noreply.github.com>
 tyronecai <tyronecai@163.com>
+Uber <20715993+YagUber@users.noreply.github.com>
 ubergarm <leimgrub@gmail.com>
 ubik2 <ubik2@users.noreply.github.com>
 UEXTM.com <84163508+uextm@users.noreply.github.com>
@@ -2242,4 +2266,5 @@ Zijun Yu <zijun.yu@intel.com>
 zql <37731799+zqlcode@users.noreply.github.com>
 zrm <trustiosity.zrm@gmail.com>
 Zsapi <martin1.zsapka@gmail.com>
+zyang-dev <267119621+zyang-dev@users.noreply.github.com>
 zzzzwc <tiddar@foxmail.com>
@@ -4,7 +4,7 @@ include(CheckIncludeFileCXX)

 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 8)
+set(LLAMA_VERSION_MINOR 9)
 set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

@@ -150,6 +150,7 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/win-app.cmake)

 if (NOT DEFINED LLAMA_BUILD_NUMBER)
    set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
@@ -269,6 +270,8 @@ if (LLAMA_BUILD_MTMD AND NOT (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS))
    add_subdirectory(tools/mtmd)
 endif()

+llama_apply_app_manifests()
+
 #
 # install
 #
@@ -87,6 +87,7 @@
 /src/llama-graph.*                      @CISC
 /src/llama-model.*                      @CISC
 /src/llama-vocab.*                      @CISC
+/src/llama-moe-cache.*                  @am17an
 /src/models/                            @ggml-org/llama-models
 /tests/                                 @ggerganov
 /tests/test-chat.*                      @pwilkin
@@ -13,7 +13,7 @@
 [![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)

-[ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Asanmai%20OR%20author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Ajhen0409%20OR%20author%3Abartowski1182%20OR%20author%3Anikwen%20OR%20author%3Ahipudding%20OR%20author%3Aravi9%20OR%20author%3AServeurpersoCom%20OR%20author%3Athelittlefireman%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3Amarty1885%20OR%20author%3A0cc4m%20OR%20author%3ATitaniumtown%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3AmasterFoad%20OR%20author%3Aallozaur%20OR%20author%3AAndreas-Krebbel%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Awine99%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Achaxu01%20OR%20author%3Alhez%20OR%20author%3AAlekseiNikiforovIBM%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming%20OR%20author%3Aynankani)%20sort%3Aupdated-desc) / [dev stats](https://github.com/ggml-org/llama.cpp-dev) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Asanmai%20OR%20author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Ajhen0409%20OR%20author%3Abartowski1182%20OR%20author%3Anikwen%20OR%20author%3Ahipudding%20OR%20author%3Aravi9%20OR%20author%3AServeurpersoCom%20OR%20author%3Athelittlefireman%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3Amarty1885%20OR%20author%3A0cc4m%20OR%20author%3ATitaniumtown%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3AmasterFoad%20OR%20author%3Aallozaur%20OR%20author%3AAndreas-Krebbel%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Awine99%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Achaxu01%20OR%20author%3Alhez%20OR%20author%3AAlekseiNikiforovIBM%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming%20OR%20author%3Aynankani)%20sort%3Aupdated-desc) / [dev stats](https://github.com/ggml-org/llama.cpp-dev) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)

 </div>

@@ -79,7 +79,7 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 | --- | --- |
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
-| [CANN](docs/build.md#cann) | Ascend NPU |
+| [CANN](docs/build.md#cann) | Huawei Ascend NPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
 | [Hexagon](docs/backend/snapdragon/README.md) | Snapdragon |
@@ -296,25 +296,6 @@ function gg_run_test_llama_archs_tensor_split {
    set +e
 }

-# test_llama_archs_models
-
-function gg_run_test_llama_archs_models {
-    cd ${SRC}
-
-    set -e
-
-    # TODO: fix and re-enable `test-llama-archs` on OpenVINO
-    # TODO: the `test-llama-archs` currently does not build on Windows, so we check if the binary exists
-    if [ -z ${GG_BUILD_OPENVINO} ] && [ -f ./build-ci-release/bin/test-llama-archs ]; then
-        rm -rf build-ci-models && mkdir -p build-ci-models
-
-        # generate the dummy models used by the model-dependent tests
-        ./build-ci-release/bin/test-llama-archs -o build-ci-models 2>&1
-    fi
-
-    set +e
-}
-
 # test_scripts

 function gg_run_test_scripts {
@@ -456,12 +437,8 @@ function gg_run_qwen3_0_6b {

    (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 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
    (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

@@ -701,7 +678,6 @@ test $ret -eq 0 && gg_run ctest_release

 test $ret -eq 0 && gg_run test_backend_ops

-test $ret -eq 0 && gg_run test_llama_archs_models
 test $ret -eq 0 && gg_run test_llama_archs_tensor_split

 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -0,0 +1,66 @@
+set(LLAMA_WIN_APP_MANIFEST "${CMAKE_CURRENT_LIST_DIR}/../common/win-app.manifest")
+
+if(WIN32)
+    if(MSVC OR CMAKE_CXX_SIMULATE_ID STREQUAL "MSVC")
+        set(LLAMA_WIN_APP_MSVC_ABI TRUE)
+    else()
+        set(LLAMA_WIN_APP_MSVC_ABI FALSE)
+    endif()
+
+    if(NOT LLAMA_WIN_APP_MSVC_ABI)
+        if(CMAKE_VERSION VERSION_LESS 3.21)
+            message(FATAL_ERROR "CMake 3.21 or newer is required to build for Windows with the GNU ABI")
+        endif()
+        enable_language(RC)
+        set(LLAMA_WIN_APP_MANIFEST_RC "${CMAKE_CURRENT_BINARY_DIR}/win-app-manifest.rc")
+        file(CONFIGURE OUTPUT "${LLAMA_WIN_APP_MANIFEST_RC}" CONTENT "1 24 \"@LLAMA_WIN_APP_MANIFEST@\"\n" @ONLY)
+        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${LLAMA_WIN_APP_MANIFEST}")
+    endif()
+endif()
+
+function(llama_add_app_manifest target)
+    if(NOT WIN32)
+        return()
+    endif()
+    get_target_property(_aliased ${target} ALIASED_TARGET)
+    if(_aliased)
+        return()
+    endif()
+    get_target_property(_type ${target} TYPE)
+    if(NOT _type STREQUAL "EXECUTABLE")
+        return()
+    endif()
+    if(LLAMA_WIN_APP_MSVC_ABI)
+        target_sources(${target} PRIVATE "${LLAMA_WIN_APP_MANIFEST}")
+    else()
+        if(NOT TARGET llama-win-app-manifest-res)
+            add_library(llama-win-app-manifest-res OBJECT ${LLAMA_WIN_APP_MANIFEST_RC})
+            set_source_files_properties(${LLAMA_WIN_APP_MANIFEST_RC} PROPERTIES OBJECT_DEPENDS "${LLAMA_WIN_APP_MANIFEST}")
+        endif()
+        target_link_libraries(${target} PRIVATE "$<TARGET_OBJECTS:llama-win-app-manifest-res>")
+    endif()
+endfunction()
+
+function(llama_apply_app_manifests_dir dir)
+    get_property(_targets DIRECTORY "${dir}" PROPERTY BUILDSYSTEM_TARGETS)
+    foreach(_target IN LISTS _targets)
+        llama_add_app_manifest(${_target})
+    endforeach()
+    get_property(_subdirs DIRECTORY "${dir}" PROPERTY SUBDIRECTORIES)
+    foreach(_subdir IN LISTS _subdirs)
+        llama_apply_app_manifests_dir("${_subdir}")
+    endforeach()
+endfunction()
+
+function(llama_apply_app_manifests)
+    if(NOT WIN32)
+        return()
+    endif()
+    get_property(_subdirs DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}" PROPERTY SUBDIRECTORIES)
+    foreach(_dir IN LISTS _subdirs)
+        get_filename_component(_name "${_dir}" NAME)
+        if(_name STREQUAL "app" OR _name STREQUAL "examples" OR _name STREQUAL "pocs" OR _name STREQUAL "tests" OR _name STREQUAL "tools")
+            llama_apply_app_manifests_dir("${_dir}")
+        endif()
+    endforeach()
+endfunction()
@@ -4,6 +4,7 @@
 #include "chat.h"
 #include "common.h"
 #include "download.h"
+#include "http.h"
 #include "json-schema-to-grammar.h"
 #include "json.h"
 #include "llama.h"
@@ -12,16 +13,6 @@
 #include "speculative.h"
 #include "preset.h"

-// fix problem with std::min and std::max
-#if defined(_WIN32)
-#define WIN32_LEAN_AND_MEAN
-#ifndef NOMINMAX
-#   define NOMINMAX
-#endif
-#include <windows.h>
-#include <shellapi.h>
-#endif
-
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
@@ -38,20 +29,6 @@
 #include <thread> // for hardware_concurrency
 #include <vector>

-#ifndef __EMSCRIPTEN__
-#ifdef __linux__
-#include <linux/limits.h>
-#elif defined(_WIN32)
-#   if !defined(PATH_MAX)
-#   define PATH_MAX MAX_PATH
-#   endif
-#elif defined(_AIX)
-#include <sys/limits.h>
-#else
-#include <sys/syslimits.h>
-#endif
-#endif
-
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083

 using json = common_json;
@@ -351,7 +328,7 @@ static bool parse_bool_value(const std::string & value) {
 static std::string get_default_local_path(const std::string & url) {
    auto f = string_split<std::string>(url, '#').front();
    f = string_split<std::string>(f, '?').front();
-    return fs_path_to_utf8(fs_get_cache_file(string_split<std::string>(f, '/').back()));
+    return fs_get_cache_file(string_split<std::string>(f, '/').back()).string();
 }

 static bool spec_types_is_default(const common_params & params) {
@@ -423,7 +400,26 @@ bool common_models_handler_is_preset_repo(const common_models_handler & handler)
    return !handler.plan.preset.url.empty();
 }

+static common_download_opts download_opts_for_url(const common_download_opts & opts, const std::string & url) {
+    common_download_opts result = opts;
+    try {
+        const auto hf    = common_http_parse_url(common_get_model_endpoint());
+        const auto parts = common_http_parse_url(url);
+
+        if (parts.scheme == hf.scheme &&
+            parts.host   == hf.host &&
+            parts.port   == hf.port &&
+            string_starts_with(result.bearer_token, "hf_")) {
+            return result;
+        }
+    } catch (const std::exception &) {
+    }
+    result.bearer_token.clear();
+    return result;
+}
+
 static std::vector<common_download_task> build_url_tasks(const common_params_model & model, common_download_opts opts) {
+    opts = download_opts_for_url(opts, model.url);
    auto parts = common_download_get_all_parts(model.url);
    std::vector<common_download_task> tasks;

@@ -512,7 +508,7 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        common_download_task task;
        task.url        = params.mmproj.url;
        task.local_path = params.mmproj.path;
-        task.opts       = opts;
+        task.opts       = download_opts_for_url(opts, task.url);
        tasks.push_back(task);
    }
    bool had_spec_url = false;
@@ -520,7 +516,7 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        common_download_task task;
        task.url        = params.speculative.draft.mparams.url;
        task.local_path = params.speculative.draft.mparams.path;
-        task.opts       = opts;
+        task.opts       = download_opts_for_url(opts, task.url);
        tasks.push_back(task);
        had_spec_url = true;
    }
@@ -734,7 +730,7 @@ static void common_params_apply_system_config(common_params & params, llama_exam
    std::vector<std::filesystem::path> paths;

 #if defined(_WIN32)
-    const std::filesystem::path program_data = common_get_path_from_env("PROGRAMDATA");
+    const std::filesystem::path program_data = common_get_env("PROGRAMDATA");
    if (!program_data.empty()) {
        paths.push_back(program_data / "llama.cpp" / "config.ini");
    }
@@ -762,7 +758,7 @@ static void common_params_apply_system_config(common_params & params, llama_exam
    common_preset_context ctx(ex);
    ctx.ignore_unknown_keys = true; // the same config file is shared by all programs
    for (const auto & path : found) {
-        LOG_INF("using config file: %s\n", fs_path_to_utf8(path).c_str());
+        LOG_INF("using config file: %s\n", path.string().c_str());
        common_preset global;
        common_presets presets = ctx.load_from_ini(path, global);
        global.apply_to_params(params);
@@ -1259,44 +1255,7 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
    return true;
 }

-#ifdef _WIN32
-struct utf8_argv {
-    std::vector<std::string> buf;
-    std::vector<char*> ptrs;
-};
-
-static utf8_argv make_utf8_argv() {
-    utf8_argv out;
-    int wargc = 0;
-    LPWSTR* wargv = CommandLineToArgvW(GetCommandLineW(), &wargc);
-    if (!wargv) return out;
-
-    out.buf.reserve(wargc);
-    for (int i = 0; i < wargc; ++i) {
-        int n = WideCharToMultiByte(CP_UTF8, WC_ERR_INVALID_CHARS, wargv[i], -1, nullptr, 0, nullptr, nullptr);
-        if (n <= 0) { out.buf.emplace_back(); continue; }
-        auto& s = out.buf.emplace_back();
-        s.resize(static_cast<size_t>(n - 1));
-        (void)WideCharToMultiByte(CP_UTF8, 0, wargv[i], -1, s.data(), n, nullptr, nullptr);
-    }
-    LocalFree(wargv);
-
-    out.ptrs.reserve(out.buf.size() + 1);
-    for (auto& s : out.buf) out.ptrs.push_back(s.data());
-    out.ptrs.push_back(nullptr);
-    return out;
-}
-#endif
-
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
-#ifdef _WIN32
-    auto utf8 = make_utf8_argv();
-    // repair argv only when it matches the process command line
-    if (static_cast<int>(utf8.buf.size()) == argc) {
-        argv = utf8.ptrs.data();
-    }
-#endif
-
    auto ctx_arg = common_params_parser_init(params, ex, print_usage);
    const common_params params_org = ctx_arg.params; // the example can modify the default params

@@ -2795,6 +2754,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.moe_cache_size = (size_t) value*1024*1024;
        }
    ).set_env("LLAMA_ARG_MOE_CACHE_MIB"));
+    add_opt(common_arg(
+        {"--moe-cache-cpu-share"}, "F",
+        string_format("share of the MoE cache misses that the CPU computes from the host experts instead of uploading them to the GPU once the cache is full, from 0 to 1 (default: %.2f)", (double) params.moe_cache_cpu_share),
+        [](common_params & params, const std::string & value) {
+            const float share = std::stof(value);
+            if (share < 0.0f || share > 1.0f) {
+                throw std::invalid_argument("invalid value");
+            }
+            params.moe_cache_cpu_share = share;
+        }
+    ).set_env("LLAMA_ARG_MOE_CACHE_CPU_SHARE"));
    add_opt(common_arg(
        {"-ncffn", "--n-cpu-ffn"}, "N",
        "keep the dense FFN weights of the first N layers in the CPU\n"
@@ -219,6 +219,23 @@ int32_t common_cpu_get_num_math() {
            }
        }
    }
+#elif defined(__riscv) && defined(__linux__)
+    std::ifstream cpuinfo("/proc/cpuinfo");
+    std::string line;
+    int32_t n_a100 = 0;
+    while (std::getline(cpuinfo, line)) {
+        std::istringstream fields(line);
+        std::string key;
+        std::string separator;
+        uint64_t marchid;
+        if (fields >> key >> separator >> std::hex >> marchid &&
+            key == "marchid" && separator == ":" && marchid == UINT64_C(0x8000000041000002)) {
+            ++n_a100;
+        }
+    }
+    if (n_a100 > 0) {
+        return n_a100;
+    }
 #elif defined(__powerpc64__) || defined(__powerpc__)
    int32_t smt_factor = 1;
    int phy_cpus = common_cpu_get_num_physical_cores();
@@ -878,27 +895,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
    return true;
 }

-#include <iostream>
-
-
 #ifdef _WIN32
-std::wstring utf8_to_wstring(const std::string & str) {
-    if (str.empty()) {
-        return std::wstring();
-    }
-
-    int size = MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0);
-
-    if (size <= 0) {
-        return std::wstring();
-    }
-
-    std::wstring wstr(size, 0);
-    MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), &wstr[0], size);
-
-    return wstr;
-}
-
 std::string wstring_to_utf8(const std::wstring & str) {
    if (str.empty()) {
        return std::string();
@@ -917,12 +914,6 @@ std::string wstring_to_utf8(const std::wstring & str) {
 }
 #endif

-// returns the path as a UTF-8 string, preserving its separators
-std::string fs_path_to_utf8(const std::filesystem::path & path) {
-    const auto value = path.u8string();
-    return std::string(value.begin(), value.end());
-}
-
 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;
@@ -946,7 +937,7 @@ void fs_write_atomic(const std::filesystem::path & path, const std::string & dat

    if (file.fail() || ec) {
        std::filesystem::remove(path_tmp, ec);
-        throw std::runtime_error("failed to write file: " + fs_path_to_utf8(path));
+        throw std::runtime_error("failed to write file: " + path.string());
    }
 }

@@ -972,20 +963,9 @@ void common_set_env(const std::string & name, const std::string & value) {
 #endif
 }

-std::filesystem::path common_get_path_from_env(const std::string & name) {
-#if defined(_WIN32)
-    const std::wstring wname = utf8_to_wstring(name);
-    const wchar_t * wvalue = _wgetenv(wname.c_str());
-    return wvalue ? std::filesystem::path(wvalue) : std::filesystem::path();
-#else
-    const char * value = std::getenv(name.c_str());
-    return value ? std::filesystem::path(value) : std::filesystem::path();
-#endif
-}
-
 #if !defined(_WIN32)
 static std::filesystem::path get_home_directory() {
-    std::filesystem::path home = common_get_path_from_env("HOME");
+    std::filesystem::path home = common_get_env("HOME");
    if (!home.empty()) {
        return home;
    }
@@ -998,19 +978,19 @@ static std::filesystem::path get_home_directory() {
 #endif

 std::filesystem::path fs_get_cache_directory() {
-    std::filesystem::path cache_directory = common_get_path_from_env("LLAMA_CACHE");
+    std::filesystem::path cache_directory = common_get_env("LLAMA_CACHE");
    if (!cache_directory.empty()) {
        return cache_directory;
    }
 #if defined(_WIN32)
-    cache_directory = common_get_path_from_env("LOCALAPPDATA");
+    cache_directory = common_get_env("LOCALAPPDATA");
    if (cache_directory.empty()) {
        throw std::runtime_error("Failed to find %LOCALAPPDATA% directory");
    }
 #elif defined(__APPLE__)
    cache_directory = get_home_directory() / "Library/Caches";
 #else
-    cache_directory = common_get_path_from_env("XDG_CACHE_HOME");
+    cache_directory = common_get_env("XDG_CACHE_HOME");
    if (cache_directory.empty()) {
        cache_directory = get_home_directory() / ".cache";
    }
@@ -1021,12 +1001,12 @@ std::filesystem::path fs_get_cache_directory() {
 std::filesystem::path fs_get_config_directory() {
    std::filesystem::path config_directory;
 #if defined(_WIN32)
-    config_directory = common_get_path_from_env("APPDATA");
+    config_directory = common_get_env("APPDATA");
    if (config_directory.empty()) {
        throw std::runtime_error("Failed to find %APPDATA% directory");
    }
 #else
-    config_directory = common_get_path_from_env("XDG_CONFIG_HOME");
+    config_directory = common_get_env("XDG_CONFIG_HOME");
    if (config_directory.empty()) {
        config_directory = get_home_directory() / ".config";
    }
@@ -1040,9 +1020,9 @@ std::filesystem::path fs_get_cache_file(const std::string & filename) {
    std::error_code ec;
    common_create_directories(cache_directory, ec);
    if (ec) {
-        throw std::runtime_error("failed to create cache directory: " + fs_path_to_utf8(cache_directory));
+        throw std::runtime_error("failed to create cache directory: " + cache_directory.string());
    }
-    return cache_directory / std::filesystem::u8path(filename);
+    return cache_directory / filename;
 }

 //
@@ -1333,6 +1313,12 @@ common_init_result::common_init_result(common_params & params, bool model_only)
        LOG_INF("%s", "decision model reads the embeddings output, enabling embedding mode\n");
    }

+    if (llama_model_classifier_free_guidance(model)) {
+        params.kv_unified   = true;
+        cparams.kv_unified  = true;
+        cparams.n_seq_max  *= 2;
+    }
+
    // embeddings need the whole batch in one ubatch, so n_batch must not be larger than n_ubatch
    // (server.cpp does this check for --embedding, but before the model is loaded)
    if (cparams.embeddings && cparams.n_batch > cparams.n_ubatch) {
@@ -1772,7 +1758,8 @@ struct llama_context_params common_context_params_to_llama(const common_params &
    cparams.type_k = params.cache_type_k;
    cparams.type_v = params.cache_type_v;

-    cparams.moe_cache_size = params.moe_cache_size;
+    cparams.moe_cache_size      = params.moe_cache_size;
+    cparams.moe_cache_cpu_share = params.moe_cache_cpu_share;

    return cparams;
 }
@@ -2264,6 +2251,16 @@ int32_t common_batch::add_embd(llama_embd embd, const llama_pos * pos, llama_seq
    return size() - 1;
 }

+int32_t common_batch::add_embd_row(int32_t row, const llama_pos * pos, llama_seq_id seq_id, bool output) {
+    token t = { LLAMA_TOKEN_NULL, { 0, 0, 0, 0 }, seq_id, output, { nullptr, 0, 0 }, { nullptr, 0, 0 }, {} };
+    t.embd_row = row;
+    for (int32_t j = 0; j < n_pos; ++j) {
+        t.pos[j] = pos[j];
+    }
+    tokens.push_back(t);
+    return size() - 1;
+}
+
 llama_batch_ext * common_batch::get_sub_batch(int32_t off, int32_t n) {
    GGML_ASSERT(batch && "common_batch was not initialized with a context");
    GGML_ASSERT(off >= 0 && n >= 0 && off + n <= size());
@@ -2284,6 +2281,12 @@ llama_batch_ext * common_batch::get_sub_batch(int32_t off, int32_t n) {
            if (t.embd.data && !llama_batch_ext_set_embd_token(res, idx, t.embd)) {
                GGML_ABORT("%s: failed to set the embedding of token %d at index %d\n", __func__, t.id, i);
            }
+        } else if (t.embd_row >= 0) {
+            idx = llama_batch_ext_add_embd_row(res, t.seq_id, t.embd_row);
+            if (idx < 0) {
+                GGML_ABORT("%s: failed to add embd row %d at index %d (error %d, n = %d)\n", __func__, t.embd_row, i, idx, n);
+            }
+            llama_batch_ext_set_pos(res, idx, t.pos.data());
        } else {
            idx = llama_batch_ext_add_embd(res, t.seq_id, t.embd);
            if (idx < 0) {
@@ -597,7 +597,8 @@ struct common_params {
    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V

-    size_t moe_cache_size = 0; // GPU cache size in bytes for the MoE experts kept in the CPU, split among the GPUs like the layers
+    size_t moe_cache_size      = 0;    // GPU cache size in bytes for the MoE experts kept in the CPU, split among the GPUs like the layers
+    float  moe_cache_cpu_share = 0.0f; // share of the MoE cache misses computed on the CPU instead of uploaded

    common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;

@@ -906,13 +907,9 @@ bool glob_match(const std::string & pattern, const std::string & str);
 //

 #ifdef _WIN32
-std::wstring utf8_to_wstring(const std::string & str);
-std::string  wstring_to_utf8(const std::wstring & str);
+std::string wstring_to_utf8(const std::wstring & str);
 #endif

-// returns the path as a UTF-8 string, preserving its separators
-std::string fs_path_to_utf8(const std::filesystem::path & path);
-
 //
 // Environment utils
 //
@@ -922,9 +919,6 @@ std::string fs_path_to_utf8(const std::filesystem::path & path);
 std::string common_get_env(const std::string & name);
 void        common_set_env(const std::string & name, const std::string & value);

-// reads a path from the environment, an unset variable gives an empty path
-std::filesystem::path common_get_path_from_env(const std::string & name);
-
 //
 // Filesystem utils
 //
@@ -1099,6 +1093,7 @@ struct common_batch {
        llama_embd   state; // non-owning view of the data passed to set_embd_state(), data == NULL if none
        std::vector<llama_seq_id> seq_ids_extra; // see add_seq()
        int32_t      decision_order = 0; // see llama_batch_ext_set_decision_order()
+        int32_t      embd_row = -1; // see add_embd_row(), -1 if none
    };

    std::vector<token> tokens; // mirror of the entries, tokens[i] describes batch index i
@@ -1141,6 +1136,10 @@ struct common_batch {
    // pos points to n_pos positions
    int32_t add_embd(llama_embd embd, const llama_pos * pos, llama_seq_id seq_id, bool output);

+    // add an entry whose embedding is row `row` of the layer input device buffer of the context's ctx_other
+    // pos points to n_pos positions
+    int32_t add_embd_row(int32_t row, const llama_pos * pos, llama_seq_id seq_id, bool output);
+
    int32_t size() const { return (int32_t) tokens.size(); }
 };

@@ -1291,6 +1290,20 @@ enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);
 // prompt utils
 //

+// leave default-inserted elements uninitialized; callers must fill them before use
+template <typename T>
+struct common_no_init_allocator : std::allocator<T> {
+    template <typename U> struct rebind { using other = common_no_init_allocator<U>; };
+
+    common_no_init_allocator() = default;
+    template <typename U> common_no_init_allocator(const common_no_init_allocator<U> &) {}
+
+    template <typename U> void construct(U * p) { ::new ((void *) p) U; }
+};
+
+// sequence state data, filled by llama_state_seq_get_data_ext()
+using common_state_data = std::vector<uint8_t, common_no_init_allocator<uint8_t>>;
+
 struct common_prompt_checkpoint {
    int64_t n_tokens;

@@ -1300,8 +1313,8 @@ struct common_prompt_checkpoint {
    llama_pos pos_min;
    llama_pos pos_max;

-    std::vector<uint8_t> data_tgt;
-    std::vector<uint8_t> data_dft;
+    common_state_data data_tgt;
+    common_state_data data_dft;

    // (optional) speculative-decoding implementation state stashed with the checkpoint
    // (e.g. eagle3's deferred-boundary g_embd row)
@@ -45,7 +45,7 @@ static const std::string PARTIAL_SUFFIX = ".downloadInProgress";

 static void write_etag(const std::string & path, const std::string & etag) {
    const std::string etag_path = path + ".etag";
-    fs_write_atomic(std::filesystem::u8path(etag_path), etag);
+    fs_write_atomic(etag_path, etag);
    LOG_DBG("%s: file etag saved: %s\n", __func__, etag_path.c_str());
 }

@@ -293,8 +293,7 @@ static int common_download_file_single_online(const std::string & url,

    // 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)));
+        return std::filesystem::exists(path) || (!alt_path.empty() && std::filesystem::exists(alt_path));
    };

    if (skip_etag && cache_hit()) {
@@ -511,7 +510,7 @@ int common_download_file_single(const std::string & url,
        return common_download_file_single_online(url, path, online_opts, skip_etag, lock_path, alt_path);
    }

-    if (!std::filesystem::exists(std::filesystem::u8path(path))) {
+    if (!std::filesystem::exists(path)) {
        LOG_ERR("%s: required file is not available in cache (offline mode): %s\n", __func__, path.c_str());
        return -1;
    }
@@ -748,7 +747,8 @@ static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
    }

    for (const auto & t : tags) {
-        std::regex pattern(t + "[.-]", std::regex::icase);
+        // match the tag as a full token, so ":F16" does not match a BF16 file
+        std::regex pattern("(^|[-./])" + t + "[.-]", std::regex::icase);
        for (const auto & f : files) {
            if (gguf_filename_is_model(f.path) &&
                std::regex_search(f.path, pattern)) {
@@ -1005,7 +1005,7 @@ std::string common_docker_resolve_model(const std::string & docker) {
        std::string model_filename = repo;
        std::replace(model_filename.begin(), model_filename.end(), '/', '_');
        model_filename += "_" + tag + ".gguf";
-        std::string local_path = fs_path_to_utf8(fs_get_cache_file(model_filename));
+        std::string local_path = fs_get_cache_file(model_filename).string();

        const std::string blob_url = url_prefix + "/blobs/" + gguf_digest;
        common_download_opts opts;
@@ -1211,7 +1211,7 @@ bool common_download_remove_sidecars(const std::string & hf_repo_with_tag, const
    for (const auto & s : sidecars) {
        const auto & f = plan.*s.file;
        if (!f.local_path.empty()) {
-            to_remove.emplace_back(std::filesystem::u8path(f.local_path));
+            to_remove.emplace_back(f.local_path);
        }
    }
    if (to_remove.empty()) {
@@ -61,6 +61,7 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
        const llama_context_params * cparams,
        std::vector<ggml_backend_dev_t> & devs,
        uint32_t & hp_ngl,
+        uint32_t & hp_ngl_skip,
        uint32_t & hp_n_ctx_train,
        uint32_t & hp_n_expert,
        ggml_log_level log_level) {
@@ -152,6 +153,8 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
    if (mparams->load_mtp) {
        hp_ngl    += llama_model_n_layer_nextn(model);
    }
+    // without MTP the NextN layers hold no weights but the model loader still counts them for n_gpu_layers
+    hp_ngl_skip    = mparams->load_mtp ? 0 : llama_model_n_layer_nextn(model);
    hp_n_ctx_train = llama_model_n_ctx_train(model);
    hp_n_expert    = llama_model_n_expert(model);

@@ -240,11 +243,12 @@ common_device_memory_data_vec common_get_device_memory_data(
        const llama_context_params * cparams,
        std::vector<ggml_backend_dev_t> & devs,
        uint32_t & hp_ngl,
+        uint32_t & hp_ngl_skip,
        uint32_t & hp_n_ctx_train,
        uint32_t & hp_n_expert,
        ggml_log_level log_level) {
    std::vector<llama_device_memory_data> impl = common_get_device_memory_data_impl(
-            path_model, mparams, cparams, devs, hp_ngl, hp_n_ctx_train, hp_n_expert, log_level);
+            path_model, mparams, cparams, devs, hp_ngl, hp_ngl_skip, hp_n_ctx_train, hp_n_expert, log_level);

    common_device_memory_data_vec ret(impl.size());
    for (size_t i = 0; i < impl.size(); i++) {
@@ -270,6 +274,7 @@ static void common_params_fit_impl(

    std::vector<ggml_backend_dev_t> devs;
    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nsk = 0; // number of skipped NextN layers, counted by the model loader but not by hp_ngl
    uint32_t hp_nct = 0; // hparams.n_ctx_train
    uint32_t hp_nex = 0; // hparams.n_expert

@@ -291,6 +296,7 @@ static void common_params_fit_impl(
        if (dmds_extra.empty() || n_ctx_extra != cparams->n_ctx) {
            std::vector<ggml_backend_dev_t> devs_extra;
            uint32_t ngl_extra = 0;
+            uint32_t nsk_extra = 0;
            uint32_t nct_extra = 0;
            uint32_t nex_extra = 0;

@@ -302,7 +308,7 @@ static void common_params_fit_impl(
            dmds_t measured;
            try {
                measured = common_get_device_memory_data_impl(
-                    extra->path_model, extra->mparams, extra->cparams, devs_extra, ngl_extra, nct_extra, nex_extra, log_level);
+                    extra->path_model, extra->mparams, extra->cparams, devs_extra, ngl_extra, nsk_extra, nct_extra, nex_extra, log_level);
            } catch (const std::runtime_error & e) {
                // the extra model is optional, fit the main model alone rather than giving up
                LOG_WRN("%s: failed to measure the memory of the extra model, fitting without it: %s\n", __func__, e.what());
@@ -342,7 +348,7 @@ static void common_params_fit_impl(
    // step 1: get data for default parameters and check whether any changes are necessary in the first place

    LOG_TRC("%s: getting device memory data for initial parameters:\n", __func__);
-    dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nsk, hp_nct, hp_nex, log_level);

    // saturate instead of overflowing, this also preserves the UINT32_MAX sentinel of n_ctx_min:
    const uint32_t n_ctx_max       = (uint32_t) std::min<uint64_t>(uint64_t(hp_nct)    * n_streams, UINT32_MAX);
@@ -354,7 +360,7 @@ static void common_params_fit_impl(
        if (n_streams > 1) {
            LOG_TRC("%s: context size unset and KV cache not unified -> using %" PRIu32 " for %" PRIu32 " sequences:\n",
                __func__, n_ctx_max, n_streams);
-            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nsk, hp_nct, hp_nex, log_level);
        }
    }
    add_extra_memory(dmds_full);
@@ -493,7 +499,7 @@ static void common_params_fit_impl(

                    int64_t sum_projected_used_min_ctx = 0;
                    cparams->n_ctx = n_ctx_min_total;
-                    dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nsk, hp_nct, hp_nex, log_level);
                    add_extra_memory(dmds_min_ctx);
                    if (nd == 0) {
                        sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
@@ -640,6 +646,20 @@ static void common_params_fit_impl(
        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl + 1);
        uint32_t il0 = hp_ngl + 1 - mparams.n_gpu_layers; // start index for tensor buft overrides

+        // the skipped NextN layers come right before the output layer, put them on the same device as the output layer
+        // so that the layers assigned by the model loader match the layers above
+        if (hp_nsk > 0 && mparams.n_gpu_layers > 0) {
+            mparams.n_gpu_layers += hp_nsk;
+            if (nd > 1) {
+                // the last devices have no layers if not even the output layer fits
+                size_t id_last = nd - 1;
+                while (ngl_per_device[id_last].n_layer == 0) {
+                    id_last--;
+                }
+                tensor_split[id_last] += hp_nsk;
+            }
+        }
+
        mparams.tensor_split = tensor_split;

        size_t itbo = 0;
@@ -675,7 +695,7 @@ static void common_params_fit_impl(
        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);

        dmds_t dmd_nl = common_get_device_memory_data_impl(
-            path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+            path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nsk, hp_nct, hp_nex, log_level);
        add_extra_memory(dmd_nl);

        LOG_TRC("%s: memory for test allocation by device:\n", func_name);
@@ -704,7 +724,7 @@ static void common_params_fit_impl(

        LOG_TRC("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
        dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
-            path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+            path_model, mparams, cparams, devs, hp_ngl, hp_nsk, hp_nct, hp_nex, log_level);
        add_extra_memory(dmds_cpu_moe);

        for (size_t id = 0; id < nd; id++) {
@@ -1185,10 +1205,11 @@ void common_fit_print(
        llama_context_params * cparams) {
    std::vector<ggml_backend_dev_t> devs;
    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nsk = 0; // number of skipped NextN layers, counted by the model loader but not by hp_ngl
    uint32_t hp_nct = 0; // hparams.n_ctx_train
    uint32_t hp_nex = 0; // hparams.n_expert

-    auto dmd = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
+    auto dmd = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nsk, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
    GGML_ASSERT(dmd.size() == devs.size() + 1);

    for (size_t id = 0; id < devs.size(); id++) {
@@ -104,6 +104,7 @@ common_device_memory_data_vec common_get_device_memory_data(
         const llama_context_params * cparams,
    std::vector<ggml_backend_dev_t> & devs,
                           uint32_t & hp_ngl,
+                           uint32_t & hp_ngl_skip,
                           uint32_t & hp_n_ctx_train,
                           uint32_t & hp_n_expert,
                     ggml_log_level   log_level);
@@ -45,7 +45,7 @@ static fs::path get_cache_directory() {
            {HOME_DIR,                fs::path(".cache") / "huggingface" / "hub"}
        };
        for (const auto & entry : entries) {
-            if (fs::path base = common_get_path_from_env(entry.var); !base.empty()) {
+            if (fs::path base = common_get_env(entry.var); !base.empty()) {
                return entry.path.empty() ? base : base / entry.path;
            }
        }
@@ -63,7 +63,7 @@ static fs::path get_cache_directory() {
 }

 std::string get_cache_path() {
-    return fs_path_to_utf8(get_cache_directory());
+    return get_cache_directory().string();
 }

 static std::string folder_name_to_repo(const std::string & folder) {
@@ -242,7 +242,7 @@ static std::string get_repo_commit(const std::string & repo_id,
                LOG_WRN("%s: skip invalid commit: %s\n", __func__, _commit.c_str());
                continue;
            }
-            const fs::path candidate = fs::u8path(_name);
+            const fs::path candidate = _name;

            if (!is_valid_subpath(refs_path, candidate)) {
                LOG_WRN("%s: skip invalid branch: %s\n", __func__, _name.c_str());
@@ -317,7 +317,7 @@ hf_files get_repo_files(const std::string & repo_id,
            file.repo_id = repo_id;
            file.path = item["path"].get<std::string>();

-            const fs::path subpath = fs::u8path(file.path);
+            const fs::path subpath = file.path;

            if (!is_valid_subpath(commit_path, subpath)) {
                LOG_WRN("%s: skip invalid path: %s\n", __func__, file.path.c_str());
@@ -340,11 +340,11 @@ hf_files get_repo_files(const std::string & repo_id,
            file.url = endpoint + repo_id + "/resolve/" + commit + "/" + file.path;

            fs::path final_path = commit_path / subpath;
-            file.final_path = fs_path_to_utf8(final_path);
+            file.final_path = final_path.string();

            if (!file.oid.empty() && !fs::exists(final_path)) {
                fs::path local_path = blobs_path / file.oid;
-                file.local_path = fs_path_to_utf8(local_path);
+                file.local_path = local_path.string();
            } else {
                file.local_path = file.final_path;
            }
@@ -411,7 +411,7 @@ hf_files get_cached_files(const std::string & repo_id) {
        if (!fs::exists(snapshots_path)) {
            continue;
        }
-        std::string _repo_id = folder_name_to_repo(fs_path_to_utf8(repo.path().filename()));
+        std::string _repo_id = folder_name_to_repo(repo.path().filename().string());

        if (!is_valid_repo_id(_repo_id)) {
            continue;
@@ -434,9 +434,8 @@ hf_files get_cached_files(const std::string & repo_id) {
            if (!path.empty()) {
                hf_file file;
                file.repo_id = _repo_id;
-                const auto generic_path = path.generic_u8string();
-                file.path = std::string(generic_path.begin(), generic_path.end());
-                file.local_path = fs_path_to_utf8(entry.path());
+                file.path = path.generic_string();
+                file.local_path = entry.path().string();
                file.final_path = file.local_path;
                files.push_back(std::move(file));
            }
@@ -450,8 +449,8 @@ std::string finalize_file(const hf_file & file, const std::function<bool()> & ke
    static std::atomic<bool> symlinks_disabled{false};

    std::error_code ec;
-    fs::path local_path = fs::u8path(file.local_path);
-    fs::path final_path = fs::u8path(file.final_path);
+    fs::path local_path = file.local_path;
+    fs::path final_path = file.final_path;

    if (local_path == final_path || fs::exists(final_path, ec)) {
        return file.final_path;
@@ -516,7 +515,7 @@ bool remove_cached_repo(const std::string & repo_id) {
    std::error_code ec;
    auto removed = fs::remove_all(repo_path, ec);
    if (ec) {
-        LOG_ERR("%s: failed to remove repo cache %s: %s\n", __func__, fs_path_to_utf8(repo_path).c_str(), ec.message().c_str());
+        LOG_ERR("%s: failed to remove repo cache %s: %s\n", __func__, repo_path.string().c_str(), ec.message().c_str());
        return false;
    }
    return removed > 0;
@@ -307,58 +307,13 @@ value binary_expression::execute_impl(context & ctx) const {
    throw std::runtime_error("Unknown operator \"" + op.value + "\" between " + left_val->type() + " and " + right_val->type());
 }

-static value try_builtin_func(context & ctx, const std::string & name, value & input, bool undef_on_missing = false) {
-    JJ_DEBUG("Trying built-in function '%s' for type %s", name.c_str(), input->type().c_str());
-    if (ctx.is_get_stats) {
-        value_t::stats_t::mark_used(input);
-        input->stats.ops.insert(name);
-    }
-    auto builtins = input->get_builtins();
-    auto it = builtins.find(name);
-    if (it != builtins.end()) {
-        JJ_DEBUG("Binding built-in '%s'", name.c_str());
-        return mk_val<value_func>(name, it->second, input);
-    }
-    if (undef_on_missing) {
-        return mk_val<value_undefined>(name);
-    }
-    throw std::runtime_error("Unknown (built-in) filter '" + name + "' for type " + input->type());
-}
-
 static value apply_filter(context & ctx, const statement_ptr & filter, value input) {
    JJ_DEBUG("Applying filter to %s", input->type().c_str());

-    auto set_filter_alias = [](auto & filter_id) {
-        if (filter_id == "count") {
-            filter_id = "length";
-        } else if (filter_id == "d") {
-            filter_id = "default";
-        } else if (filter_id == "e") {
-            filter_id = "escape";
-        } else if (filter_id == "trim") {
-            filter_id = "strip";
-        }
-    };
-
    if (is_stmt<identifier>(filter)) {
        auto filter_id = cast_stmt<identifier>(filter)->val;
-
-        set_filter_alias(filter_id);
        JJ_DEBUG("Applying filter '%s' to %s", filter_id.c_str(), input->type().c_str());
-        // TODO: Refactor filters so this coercion can be done automatically
-        if (!input->is_undefined() && !is_val<value_string>(input) && (
-            filter_id == "capitalize" ||
-            filter_id == "lower" ||
-            filter_id == "replace" ||
-            filter_id == "strip" ||
-            filter_id == "title" ||
-            filter_id == "upper" ||
-            filter_id == "wordcount"
-        )) {
-            JJ_DEBUG("Coercing %s to String for '%s' filter", input->type().c_str(), filter_id.c_str());
-            input = mk_val<value_string>(input->as_string());
-        }
-        return try_builtin_func(ctx, filter_id, input)->invoke(func_args(ctx));
+        return apply_filter_by_name(ctx, filter_id, input, func_args(ctx));

    } else if (is_stmt<call_expression>(filter)) {
        auto call = cast_stmt<call_expression>(filter);
@@ -367,14 +322,13 @@ static value apply_filter(context & ctx, const statement_ptr & filter, value inp
        }
        auto filter_id = cast_stmt<identifier>(call->callee)->val;

-        set_filter_alias(filter_id);
        JJ_DEBUG("Applying filter '%s' with arguments to %s", filter_id.c_str(), input->type().c_str());
        func_args args(ctx);
        for (const auto & arg_expr : call->args) {
            args.push_back(arg_expr->execute(ctx));
        }

-        return try_builtin_func(ctx, filter_id, input)->invoke(args);
+        return apply_filter_by_name(ctx, filter_id, input, args);

    } else {
        throw std::runtime_error("Invalid filter expression");
@@ -72,6 +72,17 @@ const std::vector<value> & func_args::get_args() const {
    return args;
 }

+void func_args::ensure_kwargs(std::initializer_list<const char *> names) const {
+    for (const auto & arg : args) {
+        if (is_val<value_kwarg>(arg)) {
+            auto * kwarg = cast_val<value_kwarg>(arg);
+            if (std::find(names.begin(), names.end(), kwarg->key) == names.end()) {
+                throw raised_exception("Function '" + func_name + "' got an unexpected keyword argument '" + kwarg->key + "'");
+            }
+        }
+    }
+}
+
 /**
 * Function that mimics Python's array slicing.
 */
@@ -134,6 +145,14 @@ static value empty_value_fn(const func_args &) {
        return mk_val<T>();
    }
 }
+
+template<typename T>
+static value empty_value_no_args_fn(const func_args & args) {
+    args.ensure_kwargs();
+    args.ensure_count(1, 1);
+    return empty_value_fn<T>(args);
+}
+
 template<typename T>
 static value test_type_fn(const func_args & args) {
    args.ensure_count(1);
@@ -594,6 +613,52 @@ const func_builtins & global_builtins() {
    return builtins;
 }

+value try_builtin_func(context & ctx, const std::string & name, value & input, bool undef_on_missing) {
+    JJ_DEBUG("Trying built-in function '%s' for type %s", name.c_str(), input->type().c_str());
+    if (ctx.is_get_stats) {
+        value_t::stats_t::mark_used(input);
+        input->stats.ops.insert(name);
+    }
+    auto builtins = input->get_builtins();
+    auto it = builtins.find(name);
+    if (it != builtins.end()) {
+        JJ_DEBUG("Binding built-in '%s'", name.c_str());
+        return mk_val<value_func>(name, it->second, input);
+    }
+    if (undef_on_missing) {
+        return mk_val<value_undefined>(name);
+    }
+    throw std::runtime_error("Unknown (built-in) filter '" + name + "' for type " + input->type());
+}
+
+value apply_filter_by_name(context & ctx, std::string filter_id, value input, const func_args & args) {
+    if (filter_id == "count") {
+        filter_id = "length";
+    } else if (filter_id == "d") {
+        filter_id = "default";
+    } else if (filter_id == "e") {
+        filter_id = "escape";
+    } else if (filter_id == "trim") {
+        filter_id = "strip";
+    }
+
+    // TODO: Refactor filters so this coercion can be done automatically
+    if (!input->is_undefined() && !is_val<value_string>(input) && (
+        filter_id == "capitalize" ||
+        filter_id == "lower" ||
+        filter_id == "replace" ||
+        filter_id == "strip" ||
+        filter_id == "title" ||
+        filter_id == "upper" ||
+        filter_id == "wordcount"
+    )) {
+        JJ_DEBUG("Coercing %s to String for '%s' filter", input->type().c_str(), filter_id.c_str());
+        input = mk_val<value_string>(input->as_string());
+    }
+
+    return try_builtin_func(ctx, filter_id, input)->invoke(args);
+}
+

 const func_builtins & value_int_t::get_builtins() const {
    static const func_builtins builtins = {
@@ -667,16 +732,22 @@ const func_builtins & value_string_t::get_builtins() const {
    static const func_builtins builtins = {
        {"default", default_value},
        {"upper", [](const func_args & args) -> value {
+            args.ensure_kwargs();
+            args.ensure_count(1, 1);
            args.ensure_vals<value_string>();
            jinja::string str = args.get_pos(0)->as_string().uppercase();
            return mk_val<value_string>(str);
        }},
        {"lower", [](const func_args & args) -> value {
+            args.ensure_kwargs();
+            args.ensure_count(1, 1);
            args.ensure_vals<value_string>();
            jinja::string str = args.get_pos(0)->as_string().lowercase();
            return mk_val<value_string>(str);
        }},
        {"strip", [](const func_args & args) -> value {
+            args.ensure_kwargs({"chars"});
+            args.ensure_count(1, 2);
            value val_input = args.get_pos(0);
            if (!is_val<value_string>(val_input)) {
                throw raised_exception("strip() first argument must be a string");
@@ -707,11 +778,15 @@ const func_builtins & value_string_t::get_builtins() const {
            }
        }},
        {"title", [](const func_args & args) -> value {
+            args.ensure_kwargs();
+            args.ensure_count(1, 1);
            args.ensure_vals<value_string>();
            jinja::string str = args.get_pos(0)->as_string().titlecase();
            return mk_val<value_string>(str);
        }},
        {"capitalize", [](const func_args & args) -> value {
+            args.ensure_kwargs();
+            args.ensure_count(1, 1);
            args.ensure_vals<value_string>();
            jinja::string str = args.get_pos(0)->as_string().capitalize();
            return mk_val<value_string>(str);
@@ -1013,6 +1088,10 @@ const func_builtins & value_bool_t::get_builtins() const {
    return builtins;
 }

+[[noreturn]] static value array_sum_not_implemented(const func_args &) {
+    throw not_implemented_exception("Array sum builtin not implemented");
+}
+
 const func_builtins & value_array_t::get_builtins() const {
    static const func_builtins builtins = {
        {"default", default_value},
@@ -1079,7 +1158,9 @@ const func_builtins & value_array_t::get_builtins() const {
        {"select", select_or_reject<false, false>},
        {"rejectattr", select_or_reject<true, true>},
        {"reject", select_or_reject<true, false>},
+        {"sum", array_sum_not_implemented},
        {"join", [](const func_args & args) -> value {
+            args.ensure_kwargs({"d", "attribute"});
            args.ensure_count(1, 3);
            if (!is_val<value_array>(args.get_pos(0))) {
                throw raised_exception("join() first argument must be an array");
@@ -1117,13 +1198,28 @@ const func_builtins & value_array_t::get_builtins() const {
        {"tojson", tojson},
        {"map", [](const func_args & args) -> value {
            args.ensure_count(2);
-            if (!is_val<value_array>(args.get_pos(0))) {
+            value val = args.get_pos(0);
+            if (!is_val<value_array>(val)) {
                throw raised_exception("map: first argument must be an array");
            }
            if (!is_val<value_kwarg>(args.get_args().at(1))) {
-                throw not_implemented_exception("map: filter-mapping not implemented");
+                value filter = args.get_pos(1);
+                if (!is_val<value_string>(filter)) {
+                    throw raised_exception("map: filter name must be a string");
+                }
+                auto out = mk_val<value_array>();
+                func_args filter_args(args.ctx);
+                const auto & all_args = args.get_args();
+                for (size_t i = 2; i < all_args.size(); ++i) {
+                    filter_args.push_back(all_args[i]);
+                }
+                const std::string filter_id = filter->as_string().str();
+                auto arr = val->as_array();
+                for (const auto & item : arr) {
+                    out->push_back(apply_filter_by_name(args.ctx, filter_id, item, filter_args));
+                }
+                return is_val<value_tuple>(val) ? mk_val<value_tuple>(out->as_array()) : out;
            }
-            value val       = args.get_pos(0);
            value attribute = args.get_kwarg_or_pos("attribute", 1);
            value default_val = args.get_kwarg("default", mk_val<value_undefined>());
            auto out = mk_val<value_array>();
@@ -1132,7 +1228,7 @@ const func_builtins & value_array_t::get_builtins() const {
                value attr_val = get_attribute(item, attribute, default_val);
                out->push_back(attr_val);
            }
-            return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(out->as_array())) : out;
+            return is_val<value_tuple>(val) ? mk_val<value_tuple>(out->as_array()) : out;
        }},
        {"append", [](const func_args & args) -> value {
            args.ensure_count(2);
@@ -1405,14 +1501,18 @@ const func_builtins & value_none_t::get_builtins() const {
 const func_builtins & value_undefined_t::get_builtins() const {
    static const func_builtins builtins = {
        {"default", default_value},
-        {"capitalize", empty_value_fn<value_string>},
+        {"capitalize", empty_value_no_args_fn<value_string>},
        {"first", empty_value_fn<value_undefined>},
        {"items", empty_value_fn<value_array>},
-        {"join", empty_value_fn<value_string>},
+        {"join", [](const func_args & args) -> value {
+            args.ensure_kwargs({"d", "attribute"});
+            args.ensure_count(1, 3);
+            return empty_value_fn<value_string>(args);
+        }},
        {"last", empty_value_fn<value_undefined>},
        {"length", empty_value_fn<value_int>},
        {"list", empty_value_fn<value_array>},
-        {"lower", empty_value_fn<value_string>},
+        {"lower", empty_value_no_args_fn<value_string>},
        {"map", empty_value_fn<value_array>},
        {"max", empty_value_fn<value_undefined>},
        {"min", empty_value_fn<value_undefined>},
@@ -1425,12 +1525,16 @@ const func_builtins & value_undefined_t::get_builtins() const {
        {"selectattr", empty_value_fn<value_array>},
        {"sort", empty_value_fn<value_array>},
        {"string", empty_value_fn<value_string>},
-        {"strip", empty_value_fn<value_string>},
+        {"strip", [](const func_args & args) -> value {
+            args.ensure_kwargs({"chars"});
+            args.ensure_count(1, 2);
+            return empty_value_fn<value_string>(args);
+        }},
        {"sum", empty_value_fn<value_int>},
-        {"title", empty_value_fn<value_string>},
+        {"title", empty_value_no_args_fn<value_string>},
        {"truncate", empty_value_fn<value_string>},
        {"unique", empty_value_fn<value_array>},
-        {"upper", empty_value_fn<value_string>},
+        {"upper", empty_value_no_args_fn<value_string>},
        {"wordcount", empty_value_fn<value_int>},
    };
    return builtins;
@@ -7,6 +7,7 @@
 #include <cmath>
 #include <cstdint>
 #include <functional>
+#include <initializer_list>
 #include <map>
 #include <memory>
 #include <set>
@@ -176,6 +177,8 @@ protected:
 //

 const func_builtins & global_builtins();
+value try_builtin_func(context & ctx, const std::string & name, value & input, bool undef_on_missing = false);
+value apply_filter_by_name(context & ctx, std::string filter_id, value input, const func_args & args);

 std::string value_to_json(const value & val, int indent = -1, const std::string_view item_sep = ", ", const std::string_view key_sep = ": ");

@@ -658,6 +661,7 @@ public:
    size_t count() const { return args.size(); }
    void push_back(const value & val);
    void push_front(const value & val);
+    void ensure_kwargs(std::initializer_list<const char *> names = {}) const;
    void ensure_count(size_t min, size_t max = 999) const {
        size_t n = args.size();
        if (n < min || n > max) {
@@ -71,6 +71,7 @@ common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &
        auto content         = p.rule("message-content", p.until("<|end|>"));
        auto channel         = p.literal("<|channel|>") + (p.literal("commentary") | p.literal("analysis"));
        auto constrain_type  = p.chars("[A-Za-z0-9_-]", 1, -1);
+        auto constraint      = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);

        // Occasionally, gpt-oss-20b will prefix channels with this commentary
        auto stray_commentary = p.optional(p.literal("<|channel|>commentary") + p.optional(p.literal(" to=assistant")));
@@ -84,7 +85,7 @@ common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &

        auto analysis = p.ref("analysis");
        auto preamble = p.rule("preamble", p.literal("<|channel|>commentary<|message|>") + p.content(content) + end);
-        auto final_msg = p.rule("final", stray_commentary + p.literal("<|channel|>final<|message|>") + p.content(content));
+        auto final_msg = p.rule("final", stray_commentary + p.literal("<|channel|>final") + constraint + p.literal("<|message|>") + p.content(content));

        // Consume any unsolicited tool calls, e.g. builtin functions
        auto unsolicited = p.rule("unsolicited", p.atomic(p.optional(channel) + p.literal(" to=") + content + end));
@@ -92,7 +93,6 @@ common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &
        auto any = p.rule("any", preamble | analysis);

        if (has_response_format) {
-            auto constraint = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);
            auto response_format = p.rule("response-format",
                p.literal("<|channel|>final") + constraint + p.literal("<|message|>") +
                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
@@ -109,7 +109,6 @@ common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &
                const auto   params   = common_chat_tool_parameters(function);

                auto func_name  = p.literal(" to=functions.") + p.tool_name(p.literal(name));
-                auto constraint = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);
                auto args       = p.tool_args(p.schema(p.json(), "tool-" + std::to_string(tool_index) + "-schema", params));

                // recipient in role header
@@ -74,16 +74,24 @@ common_chat_params common_chat_params_init_qwen3_coder(const common_chat_templat
    data.parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        auto generation_prompt = p.literal(GEN_PREFIX);

+        // The template strips the newline before </think>, so leave it out of the reasoning
+        auto think_body  = p.until_one_of({ "\n</think>", "</think>", "<tool_call>" });
+        auto think_close = p.literal("\n</think>") | p.literal("</think>") | p.peek(p.literal("<tool_call>"));
+
+        auto reasoning_extracted = p.optional("<think>" + p.space() + p.reasoning(think_body) + think_close);
+        auto reasoning_skipped   = p.optional("<think>" + p.space() + think_body + think_close);
+
        auto reasoning = p.eps();
        if (supports_reasoning && extract_reasoning) {
-            // The template strips the newline before </think>, so leave it out of the reasoning
-            reasoning = p.optional("<think>" + p.space() +
-                                   p.reasoning(p.until_one_of({ "\n</think>", "</think>", "<tool_call>" })) +
-                                   (p.literal("\n</think>") | p.literal("</think>") | p.peek(p.literal("<tool_call>"))));
+            reasoning = reasoning_extracted;
        }

        // Response format parser
        if (has_response_format) {
+            // The answer must be bare JSON, so the think block is consumed even when it is not extracted
+            if (supports_reasoning && !extract_reasoning) {
+                reasoning = reasoning_skipped;
+            }
            return generation_prompt + (reasoning << p.content(p.schema(p.json(), "response-format", inputs.json_schema)));
        }

@@ -171,12 +171,12 @@ static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_
    std::map<std::string, std::map<std::string, std::string>> parsed;

    if (!std::filesystem::exists(path)) {
-        throw std::runtime_error("preset file does not exist: " + fs_path_to_utf8(path));
+        throw std::runtime_error("preset file does not exist: " + path.string());
    }

    std::ifstream file(path);
    if (!file.good()) {
-        throw std::runtime_error("failed to open server preset file: " + fs_path_to_utf8(path));
+        throw std::runtime_error("failed to open server preset file: " + path.string());
    }

    std::string contents((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
@@ -225,7 +225,7 @@ static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_
    common_peg_parse_context ctx(contents);
    const auto result = parser.parse(ctx);
    if (!result.success()) {
-        throw std::runtime_error("failed to parse server config file: " + fs_path_to_utf8(path));
+        throw std::runtime_error("failed to parse server config file: " + path.string());
    }

    std::string current_section = COMMON_PRESET_DEFAULT_NAME;
@@ -323,7 +323,7 @@ common_presets common_preset_context::load_from_ini(const std::filesystem::path
                }
                LOG_DBG("accepted option: %s = %s\n", key.c_str(), preset.options[opt].c_str());
            } else if (ignore_unknown_keys) {
-                LOG_WRN("ignoring option '%s' from %s: not supported by this program\n", key.c_str(), fs_path_to_utf8(path).c_str());
+                LOG_WRN("ignoring option '%s' from %s: not supported by this program\n", key.c_str(), path.string().c_str());
            } else {
                throw std::runtime_error(string_format(
                    "option '%s' not recognized in preset '%s'",
@@ -385,7 +385,7 @@ static bool is_draft_file(const std::string & fname) {
 }

 common_presets common_preset_context::load_from_models_dir(const std::string & models_dir) const {
-    const std::filesystem::path dir = std::filesystem::u8path(models_dir);
+    const std::filesystem::path dir = models_dir;
    if (!std::filesystem::exists(dir) || !std::filesystem::is_directory(dir)) {
        throw std::runtime_error(string_format("error: '%s' does not exist or is not a directory\n", models_dir.c_str()));
    }
@@ -401,7 +401,7 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
            if (!entry.is_regular_file(ec)) {
                continue;
            }
-            const std::string fname = fs_path_to_utf8(entry.path().filename());
+            const std::string fname = entry.path().filename().string();
            if (!string_ends_with(fname, ".gguf")) {
                continue;
            }
@@ -422,9 +422,9 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
        if (!path.empty()) {
            models.push_back({
                /* name        */ name,
-                /* path        */ fs_path_to_utf8(path),
-                /* path_mmproj */ fs_path_to_utf8(mmproj_file), // can be empty
-                /* path_draft  */ fs_path_to_utf8(draft_file)    // can be empty
+                /* path        */ path.string(),
+                /* path_mmproj */ mmproj_file.string(), // can be empty
+                /* path_draft  */ draft_file.string()   // can be empty
            });
        }
    };
@@ -432,13 +432,13 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
    for (const auto & entry : std::filesystem::directory_iterator(dir)) {
        std::error_code ec;
        if (entry.is_directory(ec)) {
-            scan_subdir(entry.path(), fs_path_to_utf8(entry.path().filename()));
+            scan_subdir(entry.path(), entry.path().filename().string());
            continue;
        }
        if (!entry.is_regular_file(ec)) {
            continue;
        }
-        const std::string fname = fs_path_to_utf8(entry.path().filename());
+        const std::string fname = entry.path().filename().string();
        if (!string_ends_with(fname, ".gguf")) {
            continue;
        }
@@ -450,7 +450,7 @@ common_presets common_preset_context::load_from_models_dir(const std::string & m
        string_replace_all(name, ".gguf", "");
        models.push_back({
            /* name        */ name,
-            /* path        */ fs_path_to_utf8(entry.path()),
+            /* path        */ entry.path().string(),
            /* path_mmproj */ "",
            /* path_draft  */ ""
        });
@@ -136,7 +136,8 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
            // track the end sequence within forced_tokens so it is also reported on DONE
            const int32_t match = ctx->end_matcher.advance(token);
            ctx->force_pos++;
-            if (ctx->force_pos >= ctx->forced_tokens.size()) {
+            // an end tag that is already in the prompt (e.g. "<think></think>") also ends forcing
+            if (match >= 0 || ctx->force_pos >= ctx->forced_tokens.size()) {
                ctx->state = REASONING_BUDGET_DONE;
                ctx->end_match = match;
                COM_TRC("%s", "forced sequence complete, done\n");
@@ -1027,6 +1027,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
    const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
    uint32_t        target_layer_ids_n = 0;

+    // inject target features from the target's device buffer instead of host copies (env LLAMA_DFLASH_DEV=0 disables)
+    bool use_dev_features = false;
+
    common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq,
            common_speculative_type type = COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)
        : common_speculative_impl(type, n_seq, params.draft.n_max)
@@ -1141,6 +1144,23 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
        }

+        // the device buffer keeps layers in ascending id order, the fc input expects target_layer_ids order
+        {
+            const char * env = getenv("LLAMA_DFLASH_DEV");
+            const bool ascending = std::adjacent_find(target_layer_ids, target_layer_ids + target_layer_ids_n,
+                    [](int32_t a, int32_t b) { return a >= b; }) == target_layer_ids + target_layer_ids_n;
+
+            use_dev_features = !is_lilicorr && ascending && (env == nullptr || atoi(env) != 0);
+            if (use_dev_features) {
+                llama_set_embeddings_layer_inp_dev(ctx_tgt, true);
+                use_dev_features = llama_supports_embd_row(ctx_dft);
+                if (!use_dev_features) {
+                    llama_set_embeddings_layer_inp_dev(ctx_tgt, false);
+                }
+            }
+            LOG_INF("%s: DFlash target features via %s\n", __func__, use_dev_features ? "device buffer" : "host copies");
+        }
+
        // DFlash2 reads its selector lattice from h_nextn and never consumes raw logits.
        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ !is_dflash2 && !is_lilicorr);
        llama_set_causal_attn(ctx_dft, causal_attn); // DFlash needs non-causal attention unless the model says otherwise
@@ -1233,6 +1253,25 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
            for (int32_t offset = 0; offset < n_rows; offset += n_ubatch) {
                const int32_t n_chunk = std::min(n_ubatch, n_rows - offset);

+                if (use_dev_features) {
+                    // the features are the target batch rows in the target's device buffer
+                    batch_inject.clear();
+                    for (int32_t i = 0; i < n_chunk; ++i) {
+                        const int32_t   row        = i_batch_beg[seq_id] + offset + i;
+                        const llama_pos p          = batch_in.tokens[row].pos[0];
+                        const llama_pos pos_arr[4] = { p, p, p, 0 };
+                        batch_inject.add_embd_row(row, pos_arr, seq_id, false);
+                    }
+
+                    const int32_t rc = llama_process(ctx_dft, LLAMA_PROCESS_TYPE_DECODE, batch_inject.get());
+                    if (rc != 0) {
+                        LOG_ERR("%s: llama_process(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
+                                __func__, rc, (int) n_chunk, (int) offset);
+                        return false;
+                    }
+                    continue;
+                }
+
                // gather target features per extract layer; the fused decode encodes and
                // injects them into the K/V cache at the target positions
                features_buf.resize((size_t) n_chunk * n_embd_enc);
@@ -1574,12 +1613,12 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
            }
        }

-        llama_set_embeddings_nextn(ctx_tgt, true, /*masked*/ false);
-        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
-
        is_mem_shared = llama_get_ctx_other(ctx_dft) == ctx_tgt;
        chain_heads   = n_mtp_layers > 1 && !is_mem_shared;

+        llama_set_embeddings_nextn(ctx_tgt, true, /*masked*/ is_mem_shared);
+        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
+
        if (chain_heads) {
            this->params.n_max = std::min(this->params.n_max, n_mtp_layers);

@@ -1733,16 +1772,28 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
            }

            const int32_t n_rows = i_batch_end[seq_id] - i_batch_beg[seq_id] + 1;
-            verify_h_rows[seq_id] = n_rows;
-            verify_h[seq_id].resize((size_t) n_rows * n_embd);
+            if (is_mem_shared && !batch_in.tokens[i_batch_end[seq_id]].output) {
+                verify_h_rows[seq_id] = 0;
+                continue;
+            }

-            for (int32_t i = 0; i < n_rows; ++i) {
-                const float * h = llama_get_embeddings_nextn_ith(ctx_tgt, i_batch_beg[seq_id] + i);
+            const bool all_rows_output = std::all_of(
+                    batch_in.tokens.begin() + i_batch_beg[seq_id],
+                    batch_in.tokens.begin() + i_batch_end[seq_id] + 1,
+                    [](const common_batch::token & token) { return token.output; });
+            const int32_t h_rows = is_mem_shared && !all_rows_output ? 1 : n_rows;
+            const int32_t h_offset = is_mem_shared && !all_rows_output ? i_batch_end[seq_id] : i_batch_beg[seq_id];
+
+            verify_h_rows[seq_id] = h_rows;
+            verify_h[seq_id].resize((size_t) h_rows * n_embd);
+
+            for (int32_t i = 0; i < h_rows; ++i) {
+                const float * h = llama_get_embeddings_nextn_ith(ctx_tgt, h_offset + i);
                std::memcpy(verify_h[seq_id].data() + (size_t) i * n_embd, h, row_bytes);
            }

            std::memcpy(pending_h[seq_id].data(),
-                    verify_h[seq_id].data() + (size_t) (n_rows - 1) * n_embd, row_bytes);
+                    verify_h[seq_id].data() + (size_t) (h_rows - 1) * n_embd, row_bytes);
        }

        return true;
@@ -1928,6 +1979,29 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
        const size_t row_bytes = (size_t) n_embd * sizeof(float);
        std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
    }
+
+    bool get_state(llama_seq_id seq_id, std::vector<uint8_t> & data) const override {
+        if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
+            return false;
+        }
+
+        const size_t row_bytes = (size_t) n_embd * sizeof(float);
+        data.resize(row_bytes);
+        std::memcpy(data.data(), pending_h[seq_id].data(), row_bytes);
+        return true;
+    }
+
+    void set_state(llama_seq_id seq_id, const std::vector<uint8_t> & data) override {
+        if (seq_id < 0 || seq_id >= (llama_seq_id) n_seq) {
+            return;
+        }
+
+        const size_t row_bytes = (size_t) n_embd * sizeof(float);
+        if (data.size() != row_bytes) {
+            return;
+        }
+        std::memcpy(pending_h[seq_id].data(), data.data(), row_bytes);
+    }
 };

 // state of self-speculation (simple implementation, not ngram-map)
@@ -0,0 +1,9 @@
+<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<assembly xmlns="urn:schemas-microsoft-com:asm.v1" manifestVersion="1.0">
+  <application xmlns="urn:schemas-microsoft-com:asm.v3">
+    <windowsSettings>
+      <activeCodePage xmlns="http://schemas.microsoft.com/SMI/2019/WindowsSettings">UTF-8</activeCodePage>
+      <longPathAware xmlns="http://schemas.microsoft.com/SMI/2016/WindowsSettings">true</longPathAware>
+    </windowsSettings>
+  </application>
+</assembly>
@@ -20,6 +20,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "NEOChatModel": "sensenova_u1",
    "AfmoeForCausalLM": "afmoe",
    "LagunaForCausalLM": "laguna",
+    "RicursiveForCausalLM": "ricursive",
    "FlintForCausalLM": "flint",
    "ApertusForCausalLM": "llama",
    "ArceeForCausalLM": "llama",
@@ -45,7 +46,9 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "ChameleonForConditionalGeneration": "chameleon",
    "ChatGLMForConditionalGeneration": "chatglm",
    "ChatGLMModel": "chatglm",
+    "ChatterboxModel": "chatterbox",
    "ClefModel": "clef",
+    "CloverMoeModelLM": "clover",
    "CodeShellForCausalLM": "codeshell",
    "CogVLMForCausalLM": "cogvlm",
    "Cohere2MoeForCausalLM": "command_r",
@@ -60,6 +63,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "DeepseekV2ForCausalLM": "deepseek",
    "DeepseekV3ForCausalLM": "deepseek",
    "DeepseekV32ForCausalLM": "deepseek",
+    "Xing4_0ForCausalLM": "xing",
    "DFlashDraftModel": "qwen",
    "DFlash2DraftModel": "qwen",
    "LiLiCorrDraftModel": "nemotron",
@@ -93,6 +97,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "FalconForCausalLM": "falcon",
    "FalconH1ForCausalLM": "falcon_h1",
    "FalconMambaForCausalLM": "mamba",
+    "FishOmniModel": "drama3",
    "GPT2LMHeadModel": "gpt2",
    "GPTBigCodeForCausalLM": "starcoder",
    "GPTNeoXForCausalLM": "gptneox",
@@ -153,6 +158,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "KORMoForCausalLM": "qwen",
    "KimiK25ForConditionalGeneration": "deepseek",
    "KimiK3ForConditionalGeneration": "kimi_k3",
+    "KimiK31ForConditionalGeneration": "kimi_k3",
    "KimiLinearForCausalLM": "kimi_linear",
    "KimiLinearModel": "kimi_linear",
    "KimiVLForConditionalGeneration": "deepseek",
@@ -176,6 +182,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Lfm3ForCausalLM": "lfm3",
    "Lfm3OmniForConditionalGeneration": "lfm3",
    "Lfm3VlForConditionalGeneration": "lfm3",
+    "LimiteForCausalLM": "limite",
    "Llama4ForCausalLM": "llama",
    "Llama4ForConditionalGeneration": "llama",
    "LlamaBidirectionalModel": "llama",
@@ -195,6 +202,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "MambaLMHeadModel": "mamba",
    "MapleForCausalLM": "maple",
    "MellumForCausalLM": "mellum",
+    "Metis16ForCausalLM": "metis",
    "MiMoV2FlashForCausalLM": "mimo",
    "MiMoV2ForCausalLM": "mimo",
    "MiniCPM3ForCausalLM": "minicpm",
@@ -212,6 +220,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Ministral3Model": "mistral3",
    "Mistral3ForConditionalGeneration": "mistral3",
    "Mistral3Model": "mistral3",
+    "MistralLarge4ForConditionalGeneration": "mistral4",
    "VoxtralRealtimeForConditionalGeneration": "mistral3",
    "MistralForCausalLM": "llama",
    "MixtralForCausalLM": "llama",
@@ -220,6 +229,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "ModernBertForSequenceClassification": "bert",
    "ModernBertModel": "bert",
    "NanbeigeForCausalLM": "nanbeige",
+    "Nemotron3_5TranscribeForConditionalGeneration": "nemotron",
    "NemotronForCausalLM": "nemotron",
    "NemotronHForCausalLM": "nemotron",
    "NemotronHPuzzleForCausalLM": "nemotron",
@@ -325,6 +335,7 @@ TEXT_MODEL_MAP: dict[str, str] = {

 MMPROJ_MODEL_MAP: dict[str, str] = {
    "AudioFlamingo3ForConditionalGeneration": "ultravox",
+    "ChatterboxModel": "chatterbox",
    "ClefModel": "clef",
    "CogVLMForCausalLM": "cogvlm",
    "Cohere2VisionForConditionalGeneration": "command_r",
@@ -339,6 +350,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "DotsOCRForCausalLM": "dotsocr",
    "EmbeddingGemma2Model": "gemma",
    "Exaone4_5_ForConditionalGeneration": "exaone",
+    "FishOmniModel": "drama3",
    "Gemma3ForConditionalGeneration": "gemma",
    "Gemma3nForConditionalGeneration": "gemma",
    "Gemma4ForConditionalGeneration": "gemma",
@@ -377,6 +389,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "MiniCPMOcrForConditionalGeneration": "minicpm",
    "Mistral3ForConditionalGeneration": "llava",
    "Mistral3Model": "llava",
+    "Nemotron3_5TranscribeForConditionalGeneration": "nemotron",
    "NemotronH_Omni_Reasoning_V3": "nemotron",
    "NemotronH_Nano_VL_V2": "nemotron",
    "MuseGlimmerForConditionalGeneration": "muse_glimmer",
@@ -16,6 +16,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
 from itertools import chain
+from math import prod
 from transformers import AutoConfig

 import numpy as np
@@ -155,6 +156,7 @@ class ModelBase:
        self.fuse_gate_up_exps = fuse_gate_up_exps
        self._gate_exp_buffer: dict[int, Tensor] = {}
        self._up_exp_buffer: dict[int, Tensor] = {}
+        self._unfusable_gate_up_layers: set[int] = set()
        self.fuse_qkv = fuse_qkv
        self._q_buffer: dict[int, Tensor] = {}
        self._k_buffer: dict[int, Tensor] = {}
@@ -165,6 +167,8 @@ class ModelBase:
        self._fusable_qkv_weight_layers: set[int] = set()
        self._fusable_qkv_bias_layers: set[int] = set()
        self.hparams = ModelBase.load_hparams(self.dir_model, self.is_mistral_format) if hparams is None else hparams
+        # source tensor name to filtered tensor name mapping.
+        self._source_tensor_names: dict[str, str | None] = {}
        self.model_tensors = self.index_tensors(remote_hf_model_id=remote_hf_model_id)
        self.metadata_override = metadata_override
        self.model_name = model_name
@@ -175,7 +179,9 @@ class ModelBase:
        self._nvfp4_layer_algo: dict[str, str | None] = {} # per-layer quant_algo, keyed by HF module path
        self._prec_a4: dict[str, bool] = {} # gguf tensor name -> can use 4-bit (A4) activations
        self._fp8_as_q8 = fp8_as_q8
+        self._preserve_fp8 = ftype == gguf.LlamaFileType.GUESSED and not fp8_as_q8
        self._fp8_dequantized: set[str] = set()
+        self._fp8_e4m3_preserved: set[str] = set()

        # Apply heuristics to figure out typical tensor encoding based on first tensor's dtype
        # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
@@ -221,21 +227,25 @@ class ModelBase:
        tensors: dict[str, Callable[[], Tensor]] = {}

        if remote_hf_model_id is not None:
-            is_safetensors = True
-
            logger.info(f"Using remote model with HuggingFace id: {remote_hf_model_id}")
            remote_tensors = gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id)
            for name, remote_tensor in remote_tensors.items():
                data_gen = lambda r=remote_tensor: LazyTorchTensor.from_remote_tensor(r)  # noqa: E731
-                if titem := self.filter_tensors((name, data_gen)):
+                titem = self.filter_tensors((name, data_gen))
+                self._source_tensor_names[name] = titem[0] if titem else None
+                if titem:
                    tname, tgen = titem
                    tensors[tname] = tgen

            return tensors

-        prefix = "model" if not self.is_mistral_format else "consolidated"
+        return self.index_local_tensors("model" if not self.is_mistral_format else "consolidated")
+
+    # index the local "<prefix>*.safetensors" files (or pytorch_model*.bin), through filter_tensors
+    def index_local_tensors(self, prefix: str) -> dict[str, Callable[[], Tensor]]:
+        tensors: dict[str, Callable[[], Tensor]] = {}
        part_names: list[str] = ModelBase.get_model_part_names(self.dir_model, prefix, ".safetensors")
-        is_safetensors: bool = len(part_names) > 0 or (not self.is_mistral_format and (self.dir_model / "model.safetensors.index.json").is_file())
+        is_safetensors: bool = len(part_names) > 0 or (not self.is_mistral_format and (self.dir_model / f"{prefix}.safetensors.index.json").is_file())
        if not is_safetensors:
            part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")

@@ -243,7 +253,7 @@ class ModelBase:
        tensor_names_from_parts: set[str] = set()

        if not self.is_mistral_format:
-            index_name = "model.safetensors" if is_safetensors else "pytorch_model.bin"
+            index_name = f"{prefix}.safetensors" if is_safetensors else "pytorch_model.bin"
            index_name += ".index.json"
            index_file = self.dir_model / index_name

@@ -288,7 +298,9 @@ class ModelBase:
                            data_gen = lambda data=data_torch: LazyTorchTensor.from_eager(data)  # noqa: E731
                        else:
                            data_gen = lambda data=data_torch: data  # noqa: E731
-                    if titem := self.filter_tensors((name, data_gen)):
+                    titem = self.filter_tensors((name, data_gen))
+                    self._source_tensor_names[name] = titem[0] if titem else None
+                    if titem:
                        tname, tgen = titem
                        tensors[tname] = tgen

@@ -326,7 +338,7 @@ class ModelBase:

    def dequant_model(self):
        # If all quantized tensors were already handled (e.g. pure NVFP4), skip
-        if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv")) for k in self.model_tensors):
+        if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv", ".input_scale", ".activation_scale", "_activation_scale", ".k_scale", ".v_scale")) for k in self.model_tensors):
            return

        tensors_to_remove: list[str] = []
@@ -350,6 +362,12 @@ class ModelBase:
            def dequant_simple(weight: Tensor, scale: Tensor, block_size: Sequence[int] | None = None) -> Tensor:
                scale = scale.float()

+                # ModelOpt block scales are pre-shaped for broadcasting, e.g. [out/128, 1, in/128, 1]
+                if block_size is None and scale.ndim == 2 * weight.ndim and all(n == 1 for n in scale.shape[1::2]):
+                    assert all(w % n == 0 for w, n in zip(weight.shape, scale.shape[0::2]))
+                    block_size = [w // n for w, n in zip(weight.shape, scale.shape[0::2])]
+                    scale = scale.reshape(tuple(scale.shape[0::2]))
+
                if block_size is not None:
                    dim_offset = scale.ndim - len(block_size)
                    for i, size in enumerate(block_size):
@@ -440,7 +458,7 @@ class ModelBase:

                return (unpacked * scale.unsqueeze(-1).float()).reshape(shape)

-            def dequant_fp8() -> None:
+            def dequant_fp8(fp8_configs: list[dict[str, Any]] | None = None) -> None:
                for name in self.model_tensors.keys():
                    if name.endswith(".weight_scale"):
                        weight_name = name.removesuffix("_scale")
@@ -449,14 +467,29 @@ class ModelBase:
                            continue
                        w = self.model_tensors[weight_name]
                        s = self.model_tensors[name]
+                        block_size = None
                        is_fp8_weight = False
-                        if self._fp8_as_q8:
-                            is_fp8_weight = w().dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
-                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), None)
+                        if self._fp8_as_q8 or fp8_configs is not None:
+                            weight = w()
+                            is_fp8_weight = weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
+                        if fp8_configs is not None:
+                            if not is_fp8_weight:
+                                raise ValueError(f"Expected remaining weight {weight_name!r} to be FP8, got {weight.dtype}")
+                            scale = s()
+                            is_channel_scale = scale.ndim >= 1 and scale.ndim <= weight.ndim and scale.shape[0] == weight.shape[0] and all(size == 1 for size in scale.shape[1:])
+                            if prod(scale.shape) != 1 and not is_channel_scale:
+                                block_configs = [c for c in fp8_configs if c.get("strategy") == "block"]
+                                if not block_configs or any(not c.get("block_structure") or c.get("group_size") is not None for c in block_configs):
+                                    raise ValueError(f"FP8 weight {weight_name!r} with scale shape {list(scale.shape)} requires a block layout")
+                                block_sizes = {tuple(c["block_structure"]) for c in block_configs}
+                                if len(block_sizes) != 1:
+                                    raise ValueError(f"FP8 weight {weight_name!r} has ambiguous block layouts")
+                                block_size = next(iter(block_sizes))
+                        self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                        tensors_to_remove.append(name)
-                        if is_fp8_weight:
+                        if self._fp8_as_q8 and is_fp8_weight:
                            self._fp8_dequantized.add(weight_name)
-                    if name.endswith((".input_scale", ".k_scale", ".v_scale")):
+                    if name.endswith((".input_scale", ".activation_scale", "_activation_scale", ".k_scale", ".v_scale")):
                        tensors_to_remove.append(name)

            if quant_method == "bitnet":
@@ -467,7 +500,7 @@ class ModelBase:
                        s = self.model_tensors[name]
                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_bitnet(w(), s())
                        tensors_to_remove.append(name)
-            elif quant_method == "fp8":
+            elif quant_method in ("fp8", "bsgemm-moe"):  # bsgemm-moe: drama-3, same block layout
                block_size = quant_config.get("weight_block_size")
                for name in self.model_tensors.keys():
                    if name.endswith("_scale_inv"):
@@ -521,29 +554,37 @@ class ModelBase:
                nvfp4_compressed_tensors = self._is_nvfp4_compressed_tensors(quant_method, quant_format, groups)

                if nvfp4_compressed_tensors:
-                    dequant_fp8()
+                    fp8_configs = [
+                        g["weights"] for g in groups.values()
+                        if isinstance(g, dict) and g.get("format") == "float-quantized"
+                        and g.get("weights", {}).get("type") == "float"
+                        and g.get("weights", {}).get("num_bits") == 8
+                    ]
+                    dequant_fp8(fp8_configs)
                elif len(groups) > 1:
                    raise NotImplementedError("Can't handle multiple config groups for compressed-tensors yet")
                elif quant_format == "float-quantized" or quant_format == "int-quantized" or quant_format == "naive-quantized":
                    weight_config = tuple(groups.values())[0]["weights"]
                    block_size = weight_config.get("block_structure", None)
                    strategy = weight_config.get("strategy")
-                    assert strategy == "channel" or strategy == "block"
-                    assert weight_config.get("group_size") is None  # didn't find a model using this yet
                    is_fp8 = (
                        quant_format == "float-quantized"
                        and weight_config.get("type") == "float"
                        and weight_config.get("num_bits") == 8
                    )
+                    assert strategy in ("channel", "block") or (is_fp8 and strategy == "tensor")
+                    assert weight_config.get("group_size") is None  # didn't find a model using this yet
                    for name in self.model_tensors.keys():
                        if name.endswith(".weight_scale"):
                            weight_name = name.removesuffix("_scale")
                            w = self.model_tensors[weight_name]
                            s = self.model_tensors[name]
-                            self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), block_size)
+                            self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                            tensors_to_remove.append(name)
                            if self._fp8_as_q8 and is_fp8:
                                self._fp8_dequantized.add(weight_name)
+                        if is_fp8 and name.endswith((".input_scale", ".activation_scale", "_activation_scale", ".k_scale", ".v_scale")):
+                            tensors_to_remove.append(name)
                elif quant_format == "pack-quantized":
                    weight_config = tuple(groups.values())[0]["weights"]
                    assert weight_config.get("strategy") == "group"
@@ -581,6 +622,148 @@ class ModelBase:
        for name, value in new_tensors.items():
            self.model_tensors[name] = value

+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        new_name = self.tensor_map.get_name(name, try_suffixes=(".weight",))
+        return (new_name,) if new_name is not None else ()
+
+    def _prepare_fp8_e4m3_tensors(self):
+        if not self._preserve_fp8:
+            return
+
+        scale_tensors: dict[str, list[tuple[int, float]] | np.ndarray] = {}
+        input_scale_tensors: dict[str, list[tuple[int, float]] | np.ndarray] = {}
+        consumed: list[str] = []
+        candidates: dict[str, tuple[str, Tensor, tuple[str, ...], torch.Size]] = {}
+        weight_dtypes: dict[str, torch.dtype] = {}
+        scaled_weights: set[str] = set()
+        expert_pattern = re.compile(r"(\.[^.]*experts)\.(\d+)\.")
+        scale_types = {
+            gguf.MODEL_TENSOR.ATTN_Q, gguf.MODEL_TENSOR.ATTN_K, gguf.MODEL_TENSOR.ATTN_V,
+            gguf.MODEL_TENSOR.ATTN_OUT, gguf.MODEL_TENSOR.ATTN_QKV, gguf.MODEL_TENSOR.ATTN_GATE,
+            gguf.MODEL_TENSOR.FFN_GATE, gguf.MODEL_TENSOR.FFN_UP, gguf.MODEL_TENSOR.FFN_DOWN,
+            gguf.MODEL_TENSOR.FFN_GATE_SHEXP, gguf.MODEL_TENSOR.FFN_UP_SHEXP, gguf.MODEL_TENSOR.FFN_DOWN_SHEXP,
+            gguf.MODEL_TENSOR.FFN_GATE_EXP, gguf.MODEL_TENSOR.FFN_UP_EXP, gguf.MODEL_TENSOR.FFN_DOWN_EXP,
+            gguf.MODEL_TENSOR.SSM_IN, gguf.MODEL_TENSOR.SSM_OUT, gguf.MODEL_TENSOR.SSM_ALPHA, gguf.MODEL_TENSOR.SSM_BETA,
+            gguf.MODEL_TENSOR.NEXTN_EH_PROJ, gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD, gguf.MODEL_TENSOR.OUTPUT,
+        }
+
+        for scale_name in list(self.model_tensors.keys()):
+            weight_name = None
+            if scale_name.endswith(".weight_scale"):
+                weight_name = scale_name.removesuffix("_scale")
+            elif scale_name.endswith("_scale_inv"):
+                weight_name = scale_name.removesuffix("_scale_inv")
+            elif scale_name.endswith(".qscale_weight"):
+                weight_name = scale_name.removesuffix("qscale_weight") + "weight"
+
+            if weight_name is None or weight_name not in self.model_tensors:
+                continue
+
+            scaled_weights.add(weight_name)
+            weight = self.model_tensors[weight_name]()
+            weight_dtypes[weight_name] = weight.dtype
+            if weight.dtype != torch.float8_e4m3fn or weight.ndim < 2:
+                continue
+
+            scale = self.model_tensors[scale_name]()
+            if prod(scale.shape) != 1:
+                continue
+
+            merged_name = expert_pattern.sub(r"\1.", weight_name)
+            new_names = self._map_fp8_weight_names(merged_name)
+            if not new_names:
+                continue
+            if any((mapped := self.tensor_map.get_type_and_name(new_name, try_suffixes=(".weight",))) is None or mapped[0] not in scale_types for new_name in new_names):
+                continue
+            bid = next((int(part) for part in new_names[0].split(".") if part.isdecimal()), None)
+            self._fp8_e4m3_preserved.add(weight_name)
+            if any(new_name.endswith("_norm.weight") or self.tensor_force_quant(weight_name, new_name, bid, weight.ndim) != gguf.GGMLQuantizationType.F8_E4M3 for new_name in new_names):
+                self._fp8_e4m3_preserved.discard(weight_name)
+                continue
+            candidates[weight_name] = (scale_name, scale, new_names, weight.shape)
+
+        expert_groups: dict[str, list[str]] = {}
+        for name in self.model_tensors:
+            if (match := expert_pattern.search(name)) and name.endswith(".weight"):
+                expert_groups.setdefault(name[:match.end(1)], []).append(name)
+        for names in expert_groups.values():
+            for name in names:
+                if name not in weight_dtypes:
+                    weight_dtypes[name] = self.model_tensors[name]().dtype
+            dtypes = {weight_dtypes[name] for name in names}
+            if torch.float8_e4m3fn not in dtypes:
+                continue
+            if dtypes == {torch.float8_e4m3fn} and not any(name in scaled_weights for name in names):
+                continue
+            # The converter merges expert projections together.
+            if not all(name in candidates for name in names):
+                for name in names:
+                    candidates.pop(name, None)
+                    self._fp8_e4m3_preserved.discard(name)
+                    if name not in scaled_weights:
+                        gen = self.model_tensors[name]
+                        self.model_tensors[name] = lambda gen=gen: gen().float()
+
+        for weight_name, (scale_name, scale, new_names, shape) in candidates.items():
+            scale = LazyTorchTensor.to_eager(scale).float().flatten()
+
+            weight_prefix = weight_name.removesuffix(".weight")
+            # Transformers fine-grained FP8 uses activation_scale while ModelOpt uses input_scale.
+            # Ref: https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/mistral3.py#L357-L361
+            input_scale_name = next((name for name in (
+                weight_prefix + ".input_scale",
+                weight_prefix + ".activation_scale",
+                weight_prefix + "_activation_scale",
+                weight_prefix + ".qscale_act",
+            ) if name in self.model_tensors), None)
+            input_scale = None
+            if input_scale_name is not None:
+                input_scale = self.model_tensors[input_scale_name]()
+                if prod(input_scale.shape) != 1:
+                    raise ValueError(f"FP8 input scale {input_scale_name!r} must be a scalar")
+                input_scale = LazyTorchTensor.to_eager(input_scale).float().flatten()
+                consumed.append(input_scale_name)
+
+            consumed.append(scale_name)
+
+            expert_match = expert_pattern.search(weight_name)
+            if expert_match:
+                expert_id = int(expert_match.group(2))
+                for new_name in new_names:
+                    target_name = new_name.replace(".weight", ".scale")
+                    entries = scale_tensors.setdefault(target_name, [])
+                    assert isinstance(entries, list)
+                    cast(list[tuple[int, float]], entries).append((expert_id, float(scale[0])))
+                    if input_scale is not None:
+                        target_name = new_name.replace(".weight", ".input_scale")
+                        entries = input_scale_tensors.setdefault(target_name, [])
+                        assert isinstance(entries, list)
+                        cast(list[tuple[int, float]], entries).append((expert_id, float(input_scale[0])))
+            else:
+                for new_name in new_names:
+                    mapped = self.tensor_map.get_type_and_name(new_name, try_suffixes=(".weight",)) if len(shape) == 3 else None
+                    is_packed_expert = mapped is not None and mapped[0] in (
+                        gguf.MODEL_TENSOR.FFN_GATE_EXP, gguf.MODEL_TENSOR.FFN_UP_EXP, gguf.MODEL_TENSOR.FFN_DOWN_EXP,
+                    )
+                    # Packed expert weights need one scale value per expert.
+                    n_scales = shape[0] if is_packed_expert else 1
+                    scale_tensors[new_name.replace(".weight", ".scale")] = np.repeat(scale.numpy(), n_scales)
+                    if input_scale is not None:
+                        input_scale_tensors[new_name.replace(".weight", ".input_scale")] = np.repeat(input_scale.numpy(), n_scales)
+
+        for name in consumed:
+            self.model_tensors.pop(name, None)
+
+        for name, values in chain(scale_tensors.items(), input_scale_tensors.items()):
+            if isinstance(values, list):
+                values.sort(key=lambda item: item[0])
+                scale = np.array([item[1] for item in values], dtype=np.float32)
+            else:
+                scale = values.astype(np.float32)
+            if not np.allclose(scale, 1.0, atol=1e-6):
+                logger.info(f"  + {name} (FP8 scale, shape [{scale.size}])")
+                self.gguf_writer.add_tensor(name, scale)
+
    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        name, gen = item
@@ -625,6 +808,26 @@ class ModelBase:
            raise ValueError(f"Can not map tensor {name!r}")
        return new_name

+    def prepare_gate_up_fusion(self) -> None:
+        self._unfusable_gate_up_layers.clear()
+        if not self.fuse_gate_up_exps or self._fp8_as_q8:
+            return
+
+        for name, gen in self.model_tensors.items():
+            merged_name = re.sub(r"(\.experts)\.\d+\.", r"\1.", name)
+            mapped = self.tensor_map.get_type_and_name(merged_name, try_suffixes=(".weight",))
+            if mapped is None:
+                continue
+            tensor_type, new_name = mapped
+            if tensor_type not in (gguf.MODEL_TENSOR.FFN_GATE_EXP, gguf.MODEL_TENSOR.FFN_UP_EXP, gguf.MODEL_TENSOR.FFN_GATE_UP_EXP) or not new_name.endswith(".weight"):
+                continue
+            if gen().dtype == torch.float8_e4m3fn:
+                bid = next(int(part) for part in new_name.split(".") if part.isdecimal())
+                self._unfusable_gate_up_layers.add(bid)
+
+        for bid in sorted(self._unfusable_gate_up_layers):
+            logger.warning(f"Skipping --fuse-gate-up-exps for layer {bid}: fusion of preserved FP8 expert projections is not supported")
+
    def prepare_qkv_fusion(self) -> None:
        self._fusable_qkv_weight_layers.clear()
        self._fusable_qkv_bias_layers.clear()
@@ -638,6 +841,7 @@ class ModelBase:
        }
        weights: dict[int, set[gguf.MODEL_TENSOR]] = {}
        biases: dict[int, set[gguf.MODEL_TENSOR]] = {}
+        fp8_layers: set[int] = set()

        for name in self.model_tensors:
            mapped = self.tensor_map.get_type_and_name(name, try_suffixes=(".weight", ".bias"))
@@ -652,12 +856,19 @@ class ModelBase:
                continue
            if new_name.endswith(".weight"):
                weights.setdefault(bid, set()).add(tensor_type)
+                if not self._fp8_as_q8 and self.model_tensors[name]().dtype == torch.float8_e4m3fn:
+                    fp8_layers.add(bid)
            elif new_name.endswith(".bias"):
                biases.setdefault(bid, set()).add(tensor_type)

        for bid, weight_types in weights.items():
            bias_types = biases.get(bid, set())
            if weight_types == qkv_types and (not bias_types or bias_types == qkv_types):
+                # NVFP4 weights use separate reordering and repacking hooks and are exported before this fusion pass.
+                # Separate FP8 scales cannot be represented by one fused QKV scale.
+                if bid in fp8_layers:
+                    logger.warning(f"Skipping --fuse-qkv for layer {bid}: fusion of preserved FP8 projections is not supported")
+                    continue
                self._fusable_qkv_weight_layers.add(bid)
                if bias_types:
                    self._fusable_qkv_bias_layers.add(bid)
@@ -846,7 +1057,8 @@ class ModelBase:
        new_name = self.map_tensor_name(name)

        # Handle gate/up expert tensor fusion if enabled
-        if self.fuse_gate_up_exps and bid is not None:
+        # Preserved FP8 gate/up scales must stay with their separate projections.
+        if self.fuse_gate_up_exps and bid is not None and bid not in self._unfusable_gate_up_layers:
            if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_GATE_EXP, bid):
                self._gate_exp_buffer[bid] = data_torch
            elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
@@ -903,8 +1115,46 @@ class ModelBase:

        return [(new_name, data_torch)]

+    def _tensor_requires_f32(self, new_name: str, bid: int | None, n_dims: int) -> bool:
+        # Conditions should closely match those in llama_model_quantize_internal in llama.cpp.
+        return n_dims <= 1 or new_name.endswith("_norm.weight") or (
+            any(
+                self.match_model_tensor_name(new_name, key, bid)
+                for key in (
+                    gguf.MODEL_TENSOR.FFN_GATE_INP,
+                    gguf.MODEL_TENSOR.FFN_GATE_INP_SHEXP,
+                    gguf.MODEL_TENSOR.POS_EMBD,
+                    gguf.MODEL_TENSOR.TOKEN_TYPES,
+                    gguf.MODEL_TENSOR.SSM_CONV1D,
+                    gguf.MODEL_TENSOR.SHORTCONV_CONV,
+                    gguf.MODEL_TENSOR.TIME_MIX_FIRST,
+                    gguf.MODEL_TENSOR.TIME_MIX_W1,
+                    gguf.MODEL_TENSOR.TIME_MIX_W2,
+                    gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
+                    gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
+                    gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
+                    gguf.MODEL_TENSOR.POSNET_NORM1,
+                    gguf.MODEL_TENSOR.POSNET_NORM2,
+                    gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
+                    gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
+                    gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
+                    gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
+                    # Kimi KDA conv weights should be F32
+                    gguf.MODEL_TENSOR.SSM_CONV1D_Q,
+                    gguf.MODEL_TENSOR.SSM_CONV1D_K,
+                    gguf.MODEL_TENSOR.SSM_CONV1D_V,
+                    # DSA indexer weights should be F32
+                    gguf.MODEL_TENSOR.INDEXER_PROJ,
+                )
+            )
+            or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
+        )
+
    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        del new_name, bid  # unused
+        if name in self._fp8_e4m3_preserved and n_dims >= 2:
+            if self._tensor_requires_f32(new_name, bid, n_dims):
+                return gguf.GGMLQuantizationType.F32
+            return gguf.GGMLQuantizationType.F8_E4M3
        # Force FP8-original tensors to Q8_0 when requested; Q8_0 is faster than F16/BF16.
        if self._fp8_as_q8 and name in self._fp8_dequantized and n_dims >= 2:
            return gguf.GGMLQuantizationType.Q8_0
@@ -1035,7 +1285,8 @@ class ModelBase:
        expert_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
        expert_input_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
        expert_shapes: dict[tuple[int, str], list[int]] = {}
-        n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=True) or 0
+        expert_names: dict[tuple[int, str], str] = {}
+        n_experts = self.find_hparam(["num_local_experts", "num_experts", "n_routed_experts"], optional=True) or 0
        consumed: list[str] = []

        for name in self.model_tensors.keys():
@@ -1046,14 +1297,17 @@ class ModelBase:
            input_scale_name = name.replace(".weight", ".input_scale")
            if scale_name not in self.model_tensors:
                continue
-            # Force eager materialization of lazy tensors
-            weight = LazyTorchTensor.to_eager(self.model_tensors[name]())
-            scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]())
+            weight = self.model_tensors[name]()
+            scale = self.model_tensors[scale_name]()

-            # Skip non-NVFP4 tensors(e.g. 1D scale, or float8 weight)
-            if scale.ndim < 2 or weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
+            # Leave FP8 weights and their scales for _prepare_fp8_e4m3_tensors.
+            if weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2) or scale.ndim < 2:
                continue
+            if weight.ndim != 2 or scale.shape[0] != weight.shape[0] or scale.shape[1] * 8 != weight.shape[1]:
+                raise ValueError(f"NVFP4 weight {name!r} has incompatible shapes: weight {list(weight.shape)}, scale {list(scale.shape)}")

+            weight = LazyTorchTensor.to_eager(weight)
+            scale = LazyTorchTensor.to_eager(scale)
            scale2 = LazyTorchTensor.to_eager(self.model_tensors.get(scale2_name, lambda: torch.tensor(1.0))())
            input_scale = LazyTorchTensor.to_eager(self.model_tensors.get(input_scale_name, lambda: torch.tensor(1.0))())

@@ -1080,6 +1334,7 @@ class ModelBase:
                    expert_scales[key] = []
                    expert_input_scales[key] = []
                    expert_shapes[key] = shape
+                    expert_names[key] = name
                expert_blocks[key].append((expert_id, raw.copy()))
                # Collect per-expert scale2 (scalar per expert)
                expert_scales[key].append((expert_id, float(scale2.float().sum())))
@@ -1088,13 +1343,13 @@ class ModelBase:

                # Flush when all experts for this (layer, proj) are collected
                if n_experts > 0 and len(expert_blocks[key]) >= n_experts:
-                    self._flush_nvfp4_experts(key, expert_blocks, expert_scales, expert_input_scales, expert_shapes, bid, proj_type)
+                    self._flush_nvfp4_experts(key, expert_blocks, expert_scales, expert_input_scales, expert_shapes, expert_names, bid, proj_type)
            else:
                self._repack_nvfp4(name, weight, scale, scale2, input_scale)

        # Flush any remaining experts (fallback if n_experts was unknown)
        for bid, proj_type in list(expert_blocks.keys()):
-            self._flush_nvfp4_experts((bid, proj_type), expert_blocks, expert_scales, expert_input_scales, expert_shapes, bid, proj_type)
+            self._flush_nvfp4_experts((bid, proj_type), expert_blocks, expert_scales, expert_input_scales, expert_shapes, expert_names, bid, proj_type)

        # Remove consumed tensors so get_tensors/modify_tensors won't see them
        for name in consumed:
@@ -1105,7 +1360,7 @@ class ModelBase:
            if name.endswith((".k_scale", ".v_scale")):
                del self.model_tensors[name]

-    def _flush_nvfp4_experts(self, key, expert_blocks, expert_scales, expert_input_scales, expert_shapes, bid, proj_type):
+    def _flush_nvfp4_experts(self, key, expert_blocks, expert_scales, expert_input_scales, expert_shapes, expert_names, bid, proj_type):
        experts = expert_blocks.pop(key)
        scales = expert_scales.pop(key)
        input_scales = expert_input_scales.pop(key)
@@ -1117,7 +1372,7 @@ class ModelBase:
        new_name = self.map_tensor_name(merged_name)
        logger.info(f"Repacked {new_name} with shape [{len(experts)}, {shape[0]}, {shape[1]}] and quantization NVFP4")
        self.gguf_writer.add_tensor(new_name, merged, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
-        self._tag_prec_a4(merged_name, new_name)
+        self._tag_prec_a4(expert_names.pop(key), new_name)

        scales.sort(key=lambda x: x[0])
        self._write_scales_tensor(new_name.replace(".weight", ".scale"), [s[1] for s in scales])
@@ -1168,18 +1423,25 @@ class ModelBase:
        # Per-tensor NVFP4 precision.
        self._nvfp4_layer_algo = {}
        if quant_layers:
-            # store all possible module paths and assert if a quantized layer is not in the model
-            modules: set[str] = set()
-            for name in self.model_tensors:
-                while name := name.rpartition(".")[0]:
-                    modules.add(name)
-
-            for layer_name, entry in quant_layers.items():
-                if not isinstance(entry, dict):
+            # quantized_layers keys are module paths in the source checkpoint, not tensor names
+            # resolve each source tensor to the closest quantized_layers parent and record the algo under its filtered name for _tag_prec_a4
+            layer_algo = {k: v.get("quant_algo") for k, v in quant_layers.items() if isinstance(v, dict)}
+            matched: set[str] = set()
+            for src_name, filtered_name in self._source_tensor_names.items():
+                module = src_name
+                while module := module.rpartition(".")[0]:
+                    if module in layer_algo:
+                        break
+                if not module:
                    continue
-                if titem := self.filter_tensors((layer_name, lambda: torch.empty(0))):
-                    assert titem[0] in modules, f"quantized_layers entry {layer_name!r} is not in the model tensors"
-                    self._nvfp4_layer_algo[titem[0]] = entry.get("quant_algo")
+
+                matched.add(module)
+                # filter_tensors may drop a tensors so we need to check for that here.
+                if filtered_name is not None:
+                    self._nvfp4_layer_algo[filtered_name.removesuffix(".weight").removesuffix(".bias")] = layer_algo[module]
+
+            unmatched = layer_algo.keys() - matched
+            assert not unmatched, f"quantized_layers entries not in the model tensors: {sorted(unmatched)}"

        # NVFP4 weights are repacked and written directly to gguf_writer.
        # This must run before dequant_model so NVFP4 tensors are removed
@@ -1209,8 +1471,10 @@ class ModelBase:
                            self.model_tensors[input_scale_name] = inverse_scale(self.model_tensors.pop(name))
            self._generate_nvfp4_tensors()

+        self._prepare_fp8_e4m3_tensors()
        self.dequant_model()

+        self.prepare_gate_up_fusion()
        self.prepare_qkv_fusion()

        # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
@@ -1224,10 +1488,21 @@ class ModelBase:
            if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
                continue

+            # FP8 transforms must only reorder values; norms and vectors use numeric F32 transforms.
+            preserve_fp8_e4m3 = self._preserve_fp8 and data_torch.dtype == torch.float8_e4m3fn and data_torch.ndim >= 2 and not name.endswith(("norm.weight", "norm_f.weight"))
+            if preserve_fp8_e4m3:
+                for new_name in self._map_fp8_weight_names(name):
+                    mapped_bid = next((int(part) for part in new_name.split(".") if part.isdecimal()), None)
+                    if self._tensor_requires_f32(new_name, mapped_bid, data_torch.ndim):
+                        preserve_fp8_e4m3 = False
+                        break
+                if preserve_fp8_e4m3:
+                    self._fp8_e4m3_preserved.add(name)
+
            old_dtype = data_torch.dtype

            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
+            if data_torch.dtype not in (torch.float16, torch.float32) and name not in self._fp8_e4m3_preserved:
                data_torch = data_torch.to(torch.float32)

            # use the first number-like part of the tensor name as the block id
@@ -1238,51 +1513,14 @@ class ModelBase:
                    break

            for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
-                # TODO: why do we squeeze here?
-                # data = data_torch.squeeze().numpy()
-                data = data_torch.numpy()
-
-                n_dims = len(data.shape)
+                n_dims = len(data_torch.shape)
                data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)

                # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
                if n_dims <= 1 or new_name.endswith("_norm.weight"):
                    data_qtype = gguf.GGMLQuantizationType.F32

-                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
-                # Some tensor types are always in float32
-                if data_qtype is False and (
-                    any(
-                        self.match_model_tensor_name(new_name, key, bid)
-                        for key in (
-                            gguf.MODEL_TENSOR.FFN_GATE_INP,
-                            gguf.MODEL_TENSOR.FFN_GATE_INP_SHEXP,
-                            gguf.MODEL_TENSOR.POS_EMBD,
-                            gguf.MODEL_TENSOR.TOKEN_TYPES,
-                            gguf.MODEL_TENSOR.SSM_CONV1D,
-                            gguf.MODEL_TENSOR.SHORTCONV_CONV,
-                            gguf.MODEL_TENSOR.TIME_MIX_FIRST,
-                            gguf.MODEL_TENSOR.TIME_MIX_W1,
-                            gguf.MODEL_TENSOR.TIME_MIX_W2,
-                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
-                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
-                            gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
-                            gguf.MODEL_TENSOR.POSNET_NORM1,
-                            gguf.MODEL_TENSOR.POSNET_NORM2,
-                            gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
-                            gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
-                            gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
-                            gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
-                            # Kimi KDA conv weights should be F32
-                            gguf.MODEL_TENSOR.SSM_CONV1D_Q,
-                            gguf.MODEL_TENSOR.SSM_CONV1D_K,
-                            gguf.MODEL_TENSOR.SSM_CONV1D_V,
-                            # DSA indexer weights should be F32
-                            gguf.MODEL_TENSOR.INDEXER_PROJ,
-                        )
-                    )
-                    or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
-                ):
+                if data_qtype is False and self._tensor_requires_f32(new_name, bid, n_dims):
                    data_qtype = gguf.GGMLQuantizationType.F32

                if data_qtype is False and any(
@@ -1320,16 +1558,22 @@ class ModelBase:
                    else:
                        raise ValueError(f"Unknown file type: {self.ftype.name}")

+                if data_torch.dtype == torch.float8_e4m3fn:
+                    data = data_torch.view(torch.uint8).numpy() if data_qtype == gguf.GGMLQuantizationType.F8_E4M3 else data_torch.float().numpy()
+                else:
+                    data = data_torch.numpy()
+
                # a chunked tensor quantizes as one chunk at a time, while it is written
                quantize = data.quantize if isinstance(data, gguf.LazyChunkedTensor) else (
                    lambda qtype, d=data: gguf.quants.quantize(d, qtype))

-                try:
-                    data = quantize(data_qtype)
-                except gguf.QuantError as e:
-                    logger.warning("%s, %s", e, "falling back to F16")
-                    data_qtype = gguf.GGMLQuantizationType.F16
-                    data = quantize(data_qtype)
+                if not (data_qtype == gguf.GGMLQuantizationType.F8_E4M3 and data.dtype == np.uint8):
+                    try:
+                        data = quantize(data_qtype)
+                    except gguf.QuantError as e:
+                        logger.warning("%s, %s", e, "falling back to F16")
+                        data_qtype = gguf.GGMLQuantizationType.F16
+                        data = quantize(data_qtype)

                shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape

@@ -1370,6 +1614,8 @@ class ModelBase:
                self.ftype = gguf.LlamaFileType.MOSTLY_NVFP4
            elif self._is_mxfp4:
                self.ftype = gguf.LlamaFileType.MOSTLY_MXFP4_MOE
+            elif self._fp8_e4m3_preserved:
+                self.ftype = gguf.LlamaFileType.MOSTLY_F8_E4M3

        # Generate parameter weight class (useful for leader boards) if not yet determined
        if self.metadata.size_label is None and total_params > 0:
@@ -1885,6 +2131,9 @@ class TextModel(ModelBase):
        if chkhsh == "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4":
            # ref: https://huggingface.co/evilfreelancer/ruGPT3XL
            res = "gpt-2"
+        if chkhsh == "df1bc34a6366466bcb36a9a1b1cc0afdb08cfe21c9250602c930f159085390a9":
+            # ref: https://huggingface.co/Lernex/Metis-1.6-Logos
+            res = "gpt-2"
        if chkhsh == "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7":
            # ref: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B
            res = "lfm2"
@@ -2672,7 +2921,13 @@ class TextModel(ModelBase):
        )

        if vocab.tokenizer_type == MistralTokenizerType.tekken:
-            self.gguf_writer.add_tokenizer_pre("tekken")
+            with open(vocab.fname_tokenizer, encoding="utf-8") as f:
+                pattern = json.load(f)["config"]["pattern"]
+            # newer tekken.json split numbers into right-aligned groups of 3 digits instead of single digits
+            if "\\p{N}{1,3}(?=(?:\\p{N}{3})*(?!\\p{N}))" in pattern:
+                self.gguf_writer.add_tokenizer_pre("tekken-num3")
+            else:
+                self.gguf_writer.add_tokenizer_pre("tekken")
            self.gguf_writer.add_token_merges(
                vocab.extract_vocab_merges_from_model()
            )
@@ -3149,7 +3404,7 @@ 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 ("NEOChatModel", "StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration"):
+    if model_type == ModelType.TEXT and arch in ("NEOChatModel", "StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration", "KimiK31ForConditionalGeneration", "Nemotron3_5TranscribeForConditionalGeneration"):
        return arch

    # if "architectures" is found in the sub-config, use that instead
@@ -0,0 +1,473 @@
+# Chatterbox (ResembleAI) turbo and multilingual v3, from the raw checkpoints of the official repos
+# the checkpoints ship no config.json, the variant is detected from the talker file and the hparams come from the reference code
+# talker: text_emb and speech_emb are fused into one [text | speech] vocab, speech token i is <|speech_i|>
+# mmproj: the s3gen sidecar, the voice and conditioning encoders, the position and speech tables, and the conds.pt voice
+
+from __future__ import annotations
+
+import json
+import math
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Callable, Iterable
+
+import torch
+import torch.nn.functional as F
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, TextModel, MmprojModel, LazyTorchTensor, gguf, logger
+from .llama import LlamaModel
+
+# checkpoint file prefixes, the files are "<prefix>.safetensors"
+TURBO_TALKER = "t3_turbo_v1"
+MTL_TALKER   = "t3_mtl23ls_v3"
+TURBO_S3GEN  = "s3gen_meanflow"
+MTL_S3GEN    = "s3gen_v3"
+VE           = "ve"
+
+# relative speech ids shared by both variants (start/stop_speech_token in the reference)
+SPEECH_BOS = 6561
+SPEECH_EOS = 6562
+
+# sampling defaults of the reference, samplers it does not use are disabled so the common defaults do not apply
+TURBO_TOP_K = 1000
+MTL_TOP_K   = 0
+TURBO_SAMPLING = {"min_p": 0.0,  "top_p": 0.95, "temp": 0.8, "penalty_repeat": 1.2}
+MTL_SAMPLING   = {"min_p": 0.05, "top_p": 1.0,  "temp": 0.8, "penalty_repeat": 1.2}
+# the reference repetition penalty covers the whole generated history
+PENALTY_LAST_N = -1
+
+
+def _has_file(dir_model: Path, prefix: str) -> bool:
+    return (dir_model / f"{prefix}.safetensors").is_file()
+
+
+def _is_turbo(dir_model: Path) -> bool:
+    if _has_file(dir_model, TURBO_TALKER):
+        return True
+    if _has_file(dir_model, MTL_TALKER):
+        return False
+    raise FileNotFoundError(f"no chatterbox talker checkpoint in {dir_model}")
+
+
+# talker hparams of the reference (GPT2_MEDIUM_CONFIG / LLAMA_520M_CONFIG_DICT, T3Config)
+TURBO_TEXT_CONFIG = {
+    "n_layer": 24, "n_embd": 1024, "n_inner": 4096, "n_head": 16, "n_ctx": 8196, "layer_norm_epsilon": 1e-05,
+    "vocab_size": 50276, "speech_vocab_size": 6563,
+}
+MTL_TEXT_CONFIG = {
+    "num_hidden_layers": 30, "hidden_size": 1024, "intermediate_size": 4096, "num_attention_heads": 16,
+    "num_key_value_heads": 16, "head_dim": 64, "max_position_embeddings": 131072, "rms_norm_eps": 1e-05,
+    "rope_theta": 500000.0,
+    "rope_scaling": {"factor": 8.0, "high_freq_factor": 4.0, "low_freq_factor": 1.0, "original_max_position_embeddings": 8192},
+    "vocab_size": 2454, "speech_vocab_size": 8194,
+}
+
+
+def _is_chatterbox_checkpoint(dir_model: Path) -> bool:
+    return _has_file(dir_model, VE) and (_has_file(dir_model, TURBO_TALKER) or _has_file(dir_model, MTL_TALKER))
+
+
+@ModelBase.register_hparams_loader(_is_chatterbox_checkpoint)
+def _load_chatterbox_hparams(dir_model: Path) -> dict[str, Any]:
+    turbo = _is_turbo(dir_model)
+    logger.info(f"gguf: detected Chatterbox {'turbo' if turbo else 'multilingual v3'} checkpoint")
+    text_config = dict(TURBO_TEXT_CONFIG if turbo else MTL_TEXT_CONFIG)
+    # the mmproj tensors are named explicitly, they do not go through the block tensor map
+    return {"architectures": ["ChatterboxModel"], "text_config": text_config, "audio_config": {"num_hidden_layers": 0}}
+
+
+@ModelBase.register("ChatterboxModel")
+class ChatterboxTalkerModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA  # multilingual; the turbo constructor switches to GPT2
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        self.is_turbo = _is_turbo(dir_model)
+        if self.is_turbo:
+            self.model_arch = gguf.MODEL_ARCH.GPT2
+        super().__init__(dir_model, *args, **kwargs)
+        self._text_embd: Tensor | None = None
+        self._speech_embd: Tensor | None = None
+        self._text_head: Tensor | None = None
+        self._speech_head: Tensor | None = None
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        return self.index_local_tensors(TURBO_TALKER if self.is_turbo else MTL_TALKER)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, _ = item
+        # the reference feeds inputs_embeds, the transformer input embeddings are unused
+        if name in ("tfmr.wte.weight", "tfmr.embed_tokens.weight"):
+            return None
+        if name.startswith(("cond_enc.", "text_pos_emb.", "speech_pos_emb.")):
+            return None
+        return super().filter_tensors(item)
+
+    def set_vocab(self):
+        if self.is_turbo:
+            self._set_vocab_turbo()
+        else:
+            self._set_vocab_mtl()
+
+    def _speech_token_names(self, n_speech: int) -> list[str]:
+        return [f"<|speech_{i}|>" for i in range(n_speech)]
+
+    def _set_vocab_turbo(self):
+        # stock GPT-2 BPE from the checkpoint dir, extended with the speech tokens
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        n_text = len(tokens)
+        n_speech = self.hparams["speech_vocab_size"]
+        speech = self._speech_token_names(n_speech)
+        tokens += speech
+        toktypes += [gguf.TokenType.CONTROL] * n_speech
+
+        with open(self.dir_model / "merges.txt", "r", encoding="utf-8") as f:
+            merges = [line.rstrip("\n") for line in f if line.strip() and not line.startswith("#version")]
+
+        self._write_vocab(tokens, toktypes, merges, tokpre, n_text)
+
+    def _write_vocab(self, tokens: list[str], toktypes: list[int], merges: list[str], tokpre: str, n_text: int):
+        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)
+        self.gguf_writer.add_token_merges(merges)
+
+        self.gguf_writer.add_bos_token_id(n_text + SPEECH_BOS)
+        self.gguf_writer.add_eos_token_id(n_text + SPEECH_EOS)
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(False)
+
+        # the reference samples the speech head only
+        self.gguf_writer.add_suppress_tokens(list(range(n_text)))
+
+    def _set_vocab_mtl(self):
+        # the reference BPE is char-level, it is re-encoded to byte-level with extra merges that rebuild each multi-byte char
+        # its lowercasing is folded into the embedding table, so the prompt needs no text preprocessing
+        with open(self.dir_model / "mtl_tokenizer.json", "r", encoding="utf-8") as f:
+            tok = json.load(f)
+
+        byte_map = gguf.vocab.bytes_to_unicode()
+
+        def enc(s: str) -> str:
+            return "".join(byte_map[b] for b in s.encode("utf-8"))
+
+        n_text = self.hparams["vocab_size"]
+        tokens: list[str] = [f"[unused_{i}]" for i in range(n_text)]
+        toktypes = [int(gguf.TokenType.UNUSED)] * n_text
+        char_merges: list[str] = []
+        for t, i in tok["model"]["vocab"].items():
+            if t == " ":
+                # the reference substitutes [SPACE] before encoding, the raw space token is never used
+                continue
+            tokens[i] = enc(t)
+            toktypes[i] = int(gguf.TokenType.NORMAL)
+            if len(t) == 1 and len(t.encode("utf-8")) > 1:
+                parts = [byte_map[b] for b in t.encode("utf-8")]
+                for k in range(1, len(parts)):
+                    char_merges.append("".join(parts[:k]) + " " + parts[k])
+        for entry in tok.get("added_tokens", []):
+            tokens[entry["id"]] = entry["content"]
+            toktypes[entry["id"]] = int(gguf.TokenType.CONTROL)
+
+        # raw spaces tokenize to [SPACE] directly
+        space_id = tok["model"]["vocab"]["[SPACE]"]
+        tokens[space_id] = enc(" ")
+        toktypes[space_id] = int(gguf.TokenType.NORMAL)
+
+        n_speech = self.hparams["speech_vocab_size"]
+        tokens += self._speech_token_names(n_speech)
+        toktypes += [int(gguf.TokenType.CONTROL)] * n_speech
+
+        # chars are atomic in the reference, they must form before any of its merges
+        merges = char_merges
+        for m in tok["model"].get("merges", []):
+            a, b = m if isinstance(m, list) else m.split(" ")
+            merges.append(enc(a) + " " + enc(b))
+
+        self._write_vocab(tokens, toktypes, merges, "default", n_text)
+
+    def _mtl_case_fold_pairs(self) -> list[tuple[int, int]]:
+        # (upper_id, lower_id) of the single-char vocab entries, for every script in the vocab
+        with open(self.dir_model / "mtl_tokenizer.json", "r", encoding="utf-8") as f:
+            vocab = json.load(f)["model"]["vocab"]
+        pairs: list[tuple[int, int]] = []
+        for t, i in vocab.items():
+            if len(t) != 1:
+                continue
+            low = t.lower()
+            if low != t and len(low) == 1 and low in vocab:
+                pairs.append((i, vocab[low]))
+        return pairs
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        sampling = TURBO_SAMPLING if self.is_turbo else MTL_SAMPLING
+        self.gguf_writer.add_sampling_top_k(TURBO_TOP_K if self.is_turbo else MTL_TOP_K)
+        self.gguf_writer.add_sampling_min_p(sampling["min_p"])
+        self.gguf_writer.add_sampling_top_p(sampling["top_p"])
+        self.gguf_writer.add_sampling_temp(sampling["temp"])
+        self.gguf_writer.add_sampling_penalty_repeat(sampling["penalty_repeat"])
+        self.gguf_writer.add_sampling_penalty_last_n(PENALTY_LAST_N)
+        if not self.is_turbo:
+            self.gguf_writer.add_classifier_free_guidance(True)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        if self.is_turbo:
+            return
+        # llama3 rope scaling baked into the rope_freqs factors tensor
+        rp = self.hparams["rope_scaling"]
+        dim = self.hparams["head_dim"]
+        base = self.hparams["rope_theta"]
+        freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        factor = rp["factor"]
+        low_freq_wavelen = rp["original_max_position_embeddings"] / rp["low_freq_factor"]
+        high_freq_wavelen = rp["original_max_position_embeddings"] / rp["high_freq_factor"]
+        rope_factors = []
+        for freq in freqs:
+            wavelen = 2 * math.pi / freq
+            if wavelen < high_freq_wavelen:
+                rope_factors.append(1)
+            elif wavelen > low_freq_wavelen:
+                rope_factors.append(factor)
+            else:
+                smooth = (rp["original_max_position_embeddings"] / wavelen - rp["low_freq_factor"]) / (rp["high_freq_factor"] - rp["low_freq_factor"])
+                rope_factors.append(1 / ((1 - smooth) / factor + smooth))
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
+
+    def _maybe_emit_fused(self) -> Iterable[tuple[str, Tensor]]:
+        if self._text_embd is not None and self._speech_embd is not None:
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD),
+                   torch.cat([self._text_embd, self._speech_embd], dim=0))
+            self._text_embd = None
+            self._speech_embd = None
+        if self._text_head is not None and self._speech_head is not None:
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT),
+                   torch.cat([self._text_head, self._speech_head], dim=0))
+            self._text_head = None
+            self._speech_head = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # generate_extra_tensors output comes back through here with its final name
+        if name.startswith("rope_freqs"):
+            yield (name, data_torch)
+            return
+        # fused [text | speech] vocab: embeddings and output head
+        if name == "text_emb.weight":
+            if not self.is_turbo:
+                # the reference lowercases its input, the uppercase rows take the lowercase ones
+                fold = list(range(data_torch.shape[0]))
+                for upper, lower in self._mtl_case_fold_pairs():
+                    fold[upper] = lower
+                data_torch = data_torch[torch.tensor(fold)]
+            self._text_embd = data_torch
+            yield from self._maybe_emit_fused()
+            return
+        if name == "speech_emb.weight":
+            self._speech_embd = data_torch
+            yield from self._maybe_emit_fused()
+            return
+        if name == "text_head.weight":
+            self._text_head = data_torch
+            yield from self._maybe_emit_fused()
+            return
+        if name == "speech_head.weight":
+            self._speech_head = data_torch
+            yield from self._maybe_emit_fused()
+            return
+        if name == "speech_head.bias":
+            # the text head has no bias, its zone is suppressed at sampling anyway
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT, suffix=".bias"), F.pad(data_torch, (self.hparams["vocab_size"], 0)))
+            return
+
+        assert name.startswith("tfmr.")
+        name = name[len("tfmr."):]
+
+        if self.is_turbo:
+            # HF GPT-2 layout: Conv1D style weights are stored transposed
+            if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight")):
+                data_torch = data_torch.transpose(1, 0)
+            yield (self.map_tensor_name(name), data_torch)
+            return
+
+        if name.endswith("q_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, self.hparams["num_attention_heads"], None)
+        if name.endswith("k_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, self.hparams["num_attention_heads"], self.hparams["num_key_value_heads"])
+        yield (self.map_tensor_name("model." + name), data_torch)
+
+
+@ModelBase.register("ChatterboxModel")
+class ChatterboxMmprojModel(MmprojModel):
+    has_vision_encoder = False
+    has_audio_encoder = True
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        self.is_turbo = _is_turbo(dir_model)
+        super().__init__(dir_model, *args, **kwargs)
+        self._wnorm_g: dict[str, Tensor] = {}
+        self._wnorm_v: dict[str, Tensor] = {}
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config.get("audio_config")
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        tensors: dict[str, Callable[[], Tensor]] = {}
+        for prefix in (TURBO_S3GEN if self.is_turbo else MTL_S3GEN, VE, TURBO_TALKER if self.is_turbo else MTL_TALKER):
+            tensors.update(self.index_local_tensors(prefix))
+        return tensors
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        # the s3gen, ve and talker checkpoints have disjoint top-level names
+        name, gen = item
+        new_name: str | None = None
+        # batchnorm bookkeeping (over the gguf name length cap) and the tokenizer dsp buffers are unused
+        if name.endswith("num_batches_tracked") or name in ("tokenizer.window", "tokenizer._mel_filters"):
+            return None
+        for src, dst in (
+            ("flow.encoder.",           "a.gen.fenc."),
+            ("flow.decoder.estimator.", "a.gen.est."),
+            ("mel2wav.",                "a.gen.hift."),
+            ("speaker_encoder.",        "a.spk."),
+            ("tokenizer.",              "a.s3tok."),
+            ("flow.",                   "a.gen.flow."),  # input_embedding, encoder_proj, spk_embed_affine_layer
+            ("lstm.",                   "a.ve.lstm."),
+            ("proj.",                   "a.ve.proj."),
+            ("cond_enc.",               "a.cenc."),
+        ):
+            if name.startswith(src):
+                new_name = dst + name[len(src):]
+                break
+        if name == "text_pos_emb.emb.weight":
+            new_name = "a.gen.t3.text_pos_emb"
+        elif name == "speech_pos_emb.emb.weight":
+            new_name = "a.gen.t3.speech_pos_emb"
+        elif name == "speech_emb.weight":
+            new_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_GEN_CODE_OUT_EMBD] + ".weight"
+        if new_name is None:
+            return None
+        return super().filter_tensors((new_name, gen))
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_file_type(self.ftype)
+
+        # speaker encoder, the front-end hparams are hard-coded in clip.cpp
+        self.gguf_writer.add_clip_has_audio_encoder(True)
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.CHATTERBOX_SPKENC)
+        self.gguf_writer.add_audio_projection_dim(80)
+        self.gguf_writer.add_audio_num_mel_bins(80)
+        self.gguf_writer.add_audio_block_count(0)
+        self.gguf_writer.add_audio_embedding_length(192)
+        self.gguf_writer.add_audio_head_count(1)
+        self.gguf_writer.add_audio_feed_forward_length(192)
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+
+        # audio generator (s3gen, chatterbox projector); the flow encoder shape
+        self.gguf_writer.add_clip_has_gen_audio_encoder(True)
+        self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.CHATTERBOX)
+        self.gguf_writer.add_gen_audio_projection_dim(80)
+        self.gguf_writer.add_gen_audio_embedding_length(512)
+        self.gguf_writer.add_gen_audio_feed_forward_length(2048)
+        self.gguf_writer.add_gen_audio_block_count(6)
+        self.gguf_writer.add_gen_audio_head_count(8)
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(1e-5)
+
+    def _fuse_weight_norm(self, name: str, data_torch: Tensor) -> tuple[str, Tensor] | None:
+        # torch weight_norm parametrization: weight = g * v / |v| over dims 1..n
+        base = name.split(".parametrizations.weight.original")[0]
+        if name.endswith("original0"):
+            self._wnorm_g[base] = data_torch
+        else:
+            self._wnorm_v[base] = data_torch
+        if base in self._wnorm_g and base in self._wnorm_v:
+            g = self._wnorm_g.pop(base)
+            v = self._wnorm_v.pop(base)
+            norm = v.float().norm(dim=tuple(range(1, v.dim())), keepdim=True)
+            return (base + ".weight", g.float() * v.float() / norm)
+        return None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if ".parametrizations.weight.original" in name:
+            fused = self._fuse_weight_norm(name, data_torch)
+            if fused is not None:
+                yield fused
+            return
+        yield (name, data_torch)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del name, bid
+        # tensors read on the host stay F32
+        if new_name.startswith(("a.ve.", "a.cenc.", "voice.", "a.gen.t3.", "a.gen.hift.m_source.")):
+            return gguf.GGMLQuantizationType.F32
+        # the conv ops have no BF16 kernels
+        if n_dims >= 2 and new_name.endswith((".weight", ".weight_v")):
+            return gguf.GGMLQuantizationType.F16
+        return gguf.GGMLQuantizationType.F32
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        conds = torch.load(self.dir_model / "conds.pt", map_location="cpu", weights_only=True)
+        t3c = conds["t3"] if isinstance(conds, dict) else conds.t3
+        genc = conds["gen"] if isinstance(conds, dict) else conds.gen
+        t3c = vars(t3c) if not isinstance(t3c, dict) else t3c
+        genc = vars(genc) if not isinstance(genc, dict) else genc
+
+        def talker_tensor(name: str) -> Tensor:
+            return LazyTorchTensor.to_eager(self.model_tensors[name]())
+
+        # the conds.pt voice: voice.default holds the talker conditioning rows, voice.default.ref_* the flow decoder reference
+        # 1D tensors are written as F32, the speech ids stay exact
+        yield ("voice.default.ref_tokens", genc["prompt_token"][0].float())
+        yield ("voice.default.ref_feat", genc["prompt_feat"][0].float())
+        yield ("voice.default.ref_spk", genc["embedding"][0].float())
+
+        spkr_w = talker_tensor("a.cenc.spkr_enc.weight").float()
+        spkr_b = talker_tensor("a.cenc.spkr_enc.bias").float()
+        spkr_row = spkr_w @ t3c["speaker_emb"][0].float() + spkr_b
+        speech_emb = talker_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_OUT_EMBD)).float()
+        cond_tokens = t3c["cond_prompt_speech_tokens"][0]
+
+        if self.is_turbo:
+            # turbo talker conditioning: [spkr, speech rows of the cond tokens]
+            yield ("voice.default", torch.cat([spkr_row[None], speech_emb[cond_tokens]], dim=0))
+            return
+
+        # multilingual talker conditioning [spkr, perceiver x32, emotion], from the reference perceiver over the cond speech tokens
+        with torch.no_grad():
+            speech_pos = talker_tensor("a.gen.t3.speech_pos_emb").float()
+            pse = speech_emb[cond_tokens] + speech_pos[: cond_tokens.shape[0]]
+
+            ln_w = talker_tensor("a.cenc.perceiver.attn.norm.weight").float()
+            ln_b = talker_tensor("a.cenc.perceiver.attn.norm.bias").float()
+            wq = talker_tensor("a.cenc.perceiver.attn.to_q.weight").float()
+            bq = talker_tensor("a.cenc.perceiver.attn.to_q.bias").float()
+            wk = talker_tensor("a.cenc.perceiver.attn.to_k.weight").float()
+            bk = talker_tensor("a.cenc.perceiver.attn.to_k.bias").float()
+            wv = talker_tensor("a.cenc.perceiver.attn.to_v.weight").float()
+            bv = talker_tensor("a.cenc.perceiver.attn.to_v.bias").float()
+            wo = talker_tensor("a.cenc.perceiver.attn.proj_out.weight").float()
+            bo = talker_tensor("a.cenc.perceiver.attn.proj_out.bias").float()
+            query = talker_tensor("a.cenc.perceiver.pre_attention_query")[0].float()
+
+            n_head = 4
+            n_e = query.shape[1]
+
+            def attn_block(x1: Tensor, x2: Tensor) -> Tensor:
+                nx1 = F.layer_norm(x1, (n_e,), ln_w, ln_b)
+                nx2 = F.layer_norm(x2, (n_e,), ln_w, ln_b)
+                q = (nx1 @ wq.T + bq).view(-1, n_head, n_e // n_head).transpose(0, 1)
+                k = (nx2 @ wk.T + bk).view(-1, n_head, n_e // n_head).transpose(0, 1)
+                v = (nx2 @ wv.T + bv).view(-1, n_head, n_e // n_head).transpose(0, 1)
+                ctx = F.scaled_dot_product_attention(q, k, v)
+                ctx = ctx.transpose(0, 1).reshape(-1, n_e)
+                return ctx @ wo.T + bo + x1
+
+            pre = attn_block(query, pse)
+            p32 = attn_block(pre, pre)
+
+            emo = talker_tensor("a.cenc.emotion_adv_fc.weight").float()
+            emo_row = emo[:, 0] * t3c["emotion_adv"].reshape(-1)[0]
+            yield ("voice.default", torch.cat([spkr_row[None], p32, emo_row[None]], dim=0))
@@ -0,0 +1,41 @@
+from __future__ import annotations
+
+from typing import Callable, Iterable, TYPE_CHECKING
+
+from .base import ModelBase, gguf
+from .qwen import _LinearAttentionVReorderBase
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+
+@ModelBase.register("CloverMoeModelLM")
+@ModelBase.example("Diffuspace/Clover-35B-A3B-Preview")
+class CloverMoeModel(_LinearAttentionVReorderBase):
+    model_arch = gguf.MODEL_ARCH.CLOVER_MOE
+    supports_mtp_export = False
+
+    def __init__(self, *args, **kwargs):
+        self.no_mtp = True
+        super().__init__(*args, **kwargs)
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_expert_shared_count(self.hparams.get("num_shared_experts", 1))
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, _ = item
+        if name.endswith((".conv_right_weight", ".bd_softmax_logit_scale", ".bd_softmax_history_bias", ".bd_softmax_history_c")):
+            return None
+
+        return super().filter_tensors(item)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith(".A_log") or (name.endswith("norm.weight") and not name.endswith("linear_attn.norm.weight")):
+            data_torch = data_torch.float()
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -236,6 +236,12 @@ class DeepseekModel(TextModel):
 class DeepseekV2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.DEEPSEEK2

+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        # MLA projections do not have runtime weight-scale support.
+        if name.endswith((".q_a_proj.weight", ".q_b_proj.weight", ".kv_a_proj_with_mqa.weight", ".kv_b_proj.weight")):
+            return ()
+        return super()._map_fp8_weight_names(name)
+
    # TODO @ngxson : remove this when we support MTP for deepseek models
    skip_mtp = True

@@ -0,0 +1,251 @@
+from __future__ import annotations
+
+from typing import Any, Callable, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, MmprojModel, gguf
+from .qwen import Qwen3_5MoeTextModel
+
+# Fish Audio Drama 3 (FishOmniModel)
+# text model: Qwen3.5-MoE backbone, samples the main token of each frame
+# mmproj: codebook tower and Fish DAC decoder (codec.pth)
+
+_TOWER_SHARED = "audio_decoder_mtp_shared.layers."
+_TOWER_RESIDUAL = "audio_decoder_mtp.layers."
+
+
+@ModelBase.register("FishOmniModel")
+@ModelBase.example("fishaudio/drama-3")
+class Drama3TextModel(Qwen3_5MoeTextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN35MOE
+    no_mtp = True  # the checkpoint has no MTP head
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if item[0].startswith("audio_decoder"):
+            return None
+        return super().filter_tensors(item)
+
+
+@ModelBase.register("FishOmniModel")
+@ModelBase.example("fishaudio/drama-3")
+class Drama3MmprojModel(MmprojModel):
+    has_vision_encoder = False
+    has_audio_encoder = False
+
+    _LAYER_MAP = {
+        "enorm.weight":                                 (gguf.MODEL_TENSOR.A_GEN_CODE_ENORM, True),
+        "hnorm.weight":                                 (gguf.MODEL_TENSOR.A_GEN_CODE_HNORM, True),
+        "eh_proj.weight":                               (gguf.MODEL_TENSOR.A_GEN_CODE_EH_PROJ, False),
+        "semantic_embedding.weight":                    (gguf.MODEL_TENSOR.A_GEN_CODE_BLK_EMBD, False),
+        "codebook_embedding.weight":                    (gguf.MODEL_TENSOR.A_GEN_CODE_BLK_EMBD, False),
+        "final_layernorm.weight":                       (gguf.MODEL_TENSOR.A_GEN_CODE_BLK_OUT_NORM, True),
+        "transformer_layer.attention_norm.weight":      (gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_NORM, True),
+        "transformer_layer.ffn_norm.weight":            (gguf.MODEL_TENSOR.A_GEN_CODE_FFN_NORM, True),
+        "transformer_layer.attention.q_norm.weight":    (gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_Q_NORM, True),
+        "transformer_layer.attention.k_norm.weight":    (gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_K_NORM, True),
+        "transformer_layer.attention.wo.weight":        (gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_OUT, False),
+        "transformer_layer.feed_forward.w1.weight":     (gguf.MODEL_TENSOR.A_GEN_CODE_FFN_GATE, False),
+        "transformer_layer.feed_forward.w3.weight":     (gguf.MODEL_TENSOR.A_GEN_CODE_FFN_UP, False),
+        "transformer_layer.feed_forward.w2.weight":     (gguf.MODEL_TENSOR.A_GEN_CODE_FFN_DOWN, False),
+    }
+
+    _heads: dict[int, Tensor]
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._heads = {}
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        cfg = dict(self.global_config["audio_decoder_config"])
+        cfg["num_hidden_layers"] = cfg["shared_enrichment_layers"] + cfg["num_codebooks"] - 1
+        return cfg
+
+    def set_gguf_parameters(self):
+        cfg = self.hparams
+        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.DRAMA3_GEN)
+        self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_gen_audio_embedding_length(cfg["dim"])
+        self.gguf_writer.add_gen_audio_feed_forward_length(cfg["intermediate_size"])
+        self.gguf_writer.add_gen_audio_block_count(cfg["num_hidden_layers"])
+        self.gguf_writer.add_gen_audio_head_count(cfg["n_head"])
+        self.gguf_writer.add_gen_audio_head_count_kv(cfg["n_local_heads"])
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(cfg["norm_eps"])
+        # other tower and codec hparams are hardcoded in clip.cpp
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name = item[0]
+        if name.startswith("audio_decoder") or name == "model.language_model.embed_tokens.weight":
+            return item
+        return None
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # conv1d/conv1d_dw kernels must be F16, ggml_conv_1d(_dw) has no BF16 path
+        if new_name.endswith(".weight") and (
+            new_name in ("a.gen.wav.dac.entry.weight", "a.gen.wav.dac.post_conv.weight")
+            or (".up.blk." in new_name and new_name.endswith(".dwconv.weight"))
+            or (".dac.blk." in new_name and (new_name.endswith(".conv1.weight") or new_name.endswith(".conv2.weight")))
+        ):
+            return gguf.GGMLQuantizationType.F16
+        # keep ConvTranspose1d kernels and folded codebook tables in F32
+        if new_name.endswith(".conv.weight") and (".up.blk." in new_name or ".dac.blk." in new_name):
+            return gguf.GGMLQuantizationType.F32
+        if new_name.startswith("a.gen.wav.quant."):
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        T = gguf.MODEL_TENSOR
+        cfg = self.hparams
+
+        if name.startswith("a.gen.wav."):
+            yield (name, data_torch)
+            return
+
+        if name == "model.language_model.embed_tokens.weight":
+            # semantic band of the backbone embeddings
+            begin = self.global_config["semantic_start_token_id"]
+            yield (self.format_tensor_name(T.A_GEN_CODE_OUT_EMBD), data_torch[begin:begin + cfg["vocab_size"]])
+            return
+
+        if name == "audio_decoder.codebook_embeddings.weight":
+            n_cb = cfg["num_codebooks"]
+            yield (self.format_tensor_name(T.A_GEN_CODE_EMBD), data_torch.reshape(n_cb, -1, data_torch.shape[-1]))
+            return
+
+        if name.startswith("audio_decoder_mtp.heads."):
+            n_heads = cfg["num_codebooks"] - 1
+            self._heads[int(name.split(".")[2])] = data_torch
+            if len(self._heads) < n_heads:
+                return
+            yield (self.format_tensor_name(T.A_GEN_CODE_HEAD), torch.stack([self._heads.pop(i) for i in range(n_heads)]))
+            return
+
+        for prefix, offset in ((_TOWER_SHARED, 0), (_TOWER_RESIDUAL, cfg["shared_enrichment_layers"])):
+            if not name.startswith(prefix):
+                continue
+            idx, key = name[len(prefix):].split(".", 1)
+            il = int(idx) + offset
+            if key == "transformer_layer.attention.wqkv.weight":
+                # rows are [query | gate | key | value]
+                n_q = cfg["n_head"] * cfg["head_dim"]
+                n_kv = cfg["n_local_heads"] * cfg["head_dim"]
+                q, gate, k, v = data_torch.split([n_q, n_q, n_kv, n_kv], dim=0)
+                yield (self.format_tensor_name(T.A_GEN_CODE_ATTN_Q, il), q)
+                yield (self.format_tensor_name(T.A_GEN_CODE_ATTN_GATE, il), gate)
+                yield (self.format_tensor_name(T.A_GEN_CODE_ATTN_K, il), k)
+                yield (self.format_tensor_name(T.A_GEN_CODE_ATTN_V, il), v)
+                return
+            tensor, zero_centered = self._LAYER_MAP[key]
+            if zero_centered:
+                data_torch = data_torch.float() + 1
+            yield (self.format_tensor_name(tensor, il), data_torch)
+            return
+
+        raise ValueError(f"unexpected tensor {name}")
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # only the DAC decoder, the encoder is for voice cloning (not supported yet)
+        sd = torch.load(self.dir_model / "codec.pth", map_location="cpu", mmap=True, weights_only=True)
+        sd = sd.get("state_dict", sd)
+        T = gguf.MODEL_TENSOR
+
+        def get(name: str) -> Tensor:
+            return sd[name].float()
+
+        def conv(prefix: str) -> Tensor:
+            # fold weight norm
+            if prefix + ".conv.weight" in sd:
+                return get(prefix + ".conv.weight")
+            g = get(prefix + ".conv.parametrizations.weight.original0")
+            v = get(prefix + ".conv.parametrizations.weight.original1")
+            return g * v / v.norm(dim=(1, 2), keepdim=True)
+
+        def quant_table(prefix: str) -> Tensor:
+            # fold out_proj into the codebook, so decoding is a row lookup
+            g = get(prefix + ".out_proj.weight_g")
+            v = get(prefix + ".out_proj.weight_v")
+            w = (g * v / v.norm(dim=(1, 2), keepdim=True)).squeeze(-1)
+            return get(prefix + ".codebook.weight") @ w.T + get(prefix + ".out_proj.bias")
+
+        def snake(alpha: Tensor) -> tuple[Tensor, Tensor]:
+            alpha = alpha.flatten()
+            return alpha, 1.0 / (alpha + 1e-9)
+
+        q = "quantizer."
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_FIRST_CB), quant_table(q + "semantic_quantizer.quantizers.0"))
+        n_rest = sum(1 for k in sd if k.startswith(q + "quantizer.quantizers.") and k.endswith(".codebook.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_REST_CB),
+               torch.stack([quant_table(f"{q}quantizer.quantizers.{i}") for i in range(n_rest)]))
+
+        # post_module uses interleaved RoPE, permute q/k rows so NEOX RoPE gives the same result
+        tfm = q + "post_module."
+        n_head, d_head = 16, 64
+        order = torch.cat([torch.arange(0, d_head, 2), torch.arange(1, d_head, 2)])
+
+        def neox_rows(w: Tensor) -> Tensor:
+            return w.reshape(n_head, d_head, -1)[:, order].reshape(w.shape)
+
+        n_tfm = sum(1 for k in sd if k.startswith(tfm + "layers.") and k.endswith(".ffn_norm.weight"))
+        for il in range(n_tfm):
+            p = f"{tfm}layers.{il}."
+            wq, wk, wv = get(p + "attention.wqkv.weight").chunk(3, dim=0)
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_ATTN_Q, il), neox_rows(wq))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_ATTN_K, il), neox_rows(wk))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_ATTN_V, il), wv)
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_ATTN_OUT, il), get(p + "attention.wo.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_ATTN_NORM, il), get(p + "attention_norm.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_ATTN_SCALE, il), get(p + "attention_layer_scale.gamma"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_FFN_NORM, il), get(p + "ffn_norm.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_FFN_GATE, il), get(p + "feed_forward.w1.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_FFN_UP, il), get(p + "feed_forward.w3.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_FFN_DOWN, il), get(p + "feed_forward.w2.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_TFM_FFN_SCALE, il), get(p + "ffn_layer_scale.gamma"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_OUTPUT_NORM), get(tfm + "norm.weight"))
+
+        # upsample: 2x (ConvTranspose1d + ConvNeXt)
+        for il in range(2):
+            p = f"{q}upsample.{il}."
+            yield (self.format_tensor_name(T.A_GEN_WAV_UP_CONV, il, suffix=".weight"), conv(p + "0"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_UP_CONV, il, suffix=".bias"), get(p + "0.conv.bias"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_UP_DWCONV, il, suffix=".weight"), conv(p + "1.dwconv"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_UP_DWCONV, il, suffix=".bias"), get(p + "1.dwconv.conv.bias"))
+            for key, tensor in (("norm", T.A_GEN_WAV_UP_NORM), ("pwconv1", T.A_GEN_WAV_UP_PW1), ("pwconv2", T.A_GEN_WAV_UP_PW2)):
+                yield (self.format_tensor_name(tensor, il, suffix=".weight"), get(f"{p}1.{key}.weight"))
+                yield (self.format_tensor_name(tensor, il, suffix=".bias"), get(f"{p}1.{key}.bias"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_UP_GAMMA, il, suffix=""), get(p + "1.gamma"))
+
+        # DAC: entry conv, 4x (Snake, ConvTranspose1d, 3 residual units), Snake, post conv
+        d = "decoder.model."
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_ENTRY, suffix=".weight"), conv(d + "0"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_ENTRY, suffix=".bias"), get(d + "0.conv.bias"))
+        for bid in range(4):
+            p = f"{d}{bid + 1}.block."
+            a, b = snake(get(p + "0.alpha"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_SNAKE, bid, suffix=".alpha"), a)
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_SNAKE, bid, suffix=".beta"), b)
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_CONV, bid, suffix=".weight"), conv(p + "1"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_CONV, bid, suffix=".bias"), get(p + "1.conv.bias"))
+            for xid in range(3):
+                r = f"{p}{xid + 2}.block."
+                for act, tensor in (("0", T.A_GEN_WAV_DAC_RES_ACT1), ("2", T.A_GEN_WAV_DAC_RES_ACT2)):
+                    a, b = snake(get(r + act + ".alpha"))
+                    name = gguf.TENSOR_NAMES[tensor].format(bid=bid, xid=xid)
+                    yield (name + ".alpha", a)
+                    yield (name + ".beta", b)
+                for cv, tensor in (("1", T.A_GEN_WAV_DAC_RES_CONV1), ("3", T.A_GEN_WAV_DAC_RES_CONV2)):
+                    name = gguf.TENSOR_NAMES[tensor].format(bid=bid, xid=xid)
+                    yield (name + ".weight", conv(r + cv))
+                    yield (name + ".bias", get(r + cv + ".conv.bias"))
+        a, b = snake(get(d + "5.alpha"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_SNAKE, suffix=".alpha"), a)
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_SNAKE, suffix=".beta"), b)
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_CONV, suffix=".weight"), conv(d + "6"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_CONV, suffix=".bias"), get(d + "6.conv.bias"))
@@ -343,3 +343,33 @@ class KimiK3Model(TextModel):
            return

        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("KimiK31ForConditionalGeneration")
+@ModelBase.example("moonshotai/Kimi-K3.1-Pro")
+class KimiK31Model(KimiK3Model):
+    """
+    Kimi-K3.1: K3 with the full-attention MLA layers replaced by gated DSA.
+
+    The tensor set is K3's plus a lightning indexer per full-attention layer,
+    and `dsa_config` adds the attention.indexer.* metadata.
+    """
+
+    model_arch = gguf.MODEL_ARCH.KIMI_K31
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # --- DSA indexer ---
+        dsa = self.hparams["dsa_config"]
+        dsa_layers = dsa["dsa_layers"]
+        # the layer lists are 1-indexed, like linear_attn_config's
+        assert sorted(dsa_layers) == sorted(self.hparams["linear_attn_config"]["full_attn_layers"]), \
+            "every full-attention layer is expected to own a DSA indexer"
+        self.gguf_writer.add_indexer_head_count(dsa["index_num_heads"])
+        self.gguf_writer.add_indexer_key_length(dsa["index_head_dim"])
+        self.gguf_writer.add_indexer_top_k(dsa["index_topk"])
+        self.gguf_writer.add_indexer_types([(il + 1) in dsa_layers
+                                            for il in range(self.hparams["num_hidden_layers"])])
+        # eps of the indexer key LayerNorm; the config declares no separate value
+        self.gguf_writer.add_layer_norm_eps(dsa.get("index_norm_eps", self.hparams["rms_norm_eps"]))
@@ -453,9 +453,6 @@ class LFM3AudioMmprojModel(MmprojModel):
    def tensor_force_quant(self, name, new_name, bid, n_dims):
        if new_name.startswith("a.conv1d.") or "conv_dw" in new_name or new_name.startswith("a.gen.wav.quant."):
            return gguf.GGMLQuantizationType.F32
-        # the codec decoder is small and its log-magnitude/phase output does not survive quantization
-        if new_name.startswith("a.gen.wav."):
-            return gguf.GGMLQuantizationType.F16
        return super().tensor_force_quant(name, new_name, bid, n_dims)

    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
@@ -0,0 +1,184 @@
+from __future__ import annotations
+
+from typing import Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, TextModel, gguf
+
+
+@ModelBase.register("LimiteForCausalLM")
+@ModelBase.example("paradigma-inc/limite-1b-violetto")
+class LimiteModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LIMITE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hp = self.hparams
+
+        expected = {
+            "pos_mode": "rope",
+            "qk_norm": "rms_pre_rope",
+            "rope_style": "interleaved_pairs_odd_lane_sign_flip",
+            "rope_per_layer": False,
+            "rms_norm_has_weight": False,
+            "rms_norm_eps_mode": "torch_finfo_default",
+            "mlp_type": "swiglu",
+            "attn_gate_applied": "per_head_before_o_proj",
+            "ve_head_slice": "first_num_key_value_heads",
+            "ve_applied_before_qk_norm": True,
+            "global_nope": True,
+            "global_window": -1,
+            "mudd_r_site": "resid",
+            "tie_word_embeddings": True,
+            "final_softcap": 0.0,
+        }
+        for key, value in expected.items():
+            if hp.get(key) != value:
+                raise NotImplementedError(f"Limite: unsupported {key}={hp.get(key)!r}, expected {value!r}")
+        if hp["softcap_logits"]["kind"] != "sigmoid":
+            raise NotImplementedError(f"Limite: unsupported softcap_logits kind {hp['softcap_logits']['kind']!r}")
+
+        self.n_head    = hp["num_attention_heads"]
+        self.n_head_kv = hp["num_key_value_heads"]
+        self.head_dim  = hp["head_dim"]
+        self.n_pairs   = hp["rope_n_pairs"]
+
+        self.mudd_taps: dict[int, list[int]] = {}
+        if hp.get("mudd"):
+            if not hp.get("mudd_mlp"):
+                raise NotImplementedError("Limite: mudd without mudd_mlp is not supported")
+            self.mudd_taps = {int(il): [int(t) for t in taps] for il, taps in hp["mudd_tap_idx"].items()}
+            assert all(len(taps) == hp["mudd_taps"] for taps in self.mudd_taps.values())
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hp = self.hparams
+
+        self.gguf_writer.add_key_length(self.head_dim)
+        self.gguf_writer.add_value_length(self.head_dim)
+
+        # torch rms_norm with eps=None uses the fp32 epsilon, even for bf16 inputs
+        self.gguf_writer.add_layer_norm_rms_eps(torch.finfo(torch.float32).eps)
+        self.gguf_writer.add_attention_scale(hp["attention_softmax_scale"])
+        self.gguf_writer.add_xsa_eps(hp["xsa_normalize_eps"] if hp.get("xsa") else 0.0)
+
+        # HF freq_i = base^(-i/(n_pairs-1)), ggml uses freq_base^(-i/n_pairs)
+        self.gguf_writer.add_rope_dimension_count(2 * self.n_pairs)
+        self.gguf_writer.add_rope_freq_base(hp["rope_base_local"] ** (self.n_pairs / (self.n_pairs - 1)))
+
+        # HF keeps keys k >= q - window, so a window of W spans W + 1 tokens
+        global_layers = set(hp["global_layers"])
+        self.gguf_writer.add_sliding_window(hp["sliding_window"] + 1)
+        self.gguf_writer.add_sliding_window_pattern([il not in global_layers for il in range(self.block_count)])
+
+        cap = hp["softcap_logits"]
+        self.gguf_writer.add_final_logit_sigmoid_capping(cap["a"], cap["b"], cap["c"])
+
+        if self.mudd_taps:
+            n_taps = hp["mudd_taps"]
+            tap_indices = [-1] * (self.block_count * n_taps)
+            for il, taps in self.mudd_taps.items():
+                tap_indices[il * n_taps:(il + 1) * n_taps] = taps
+            self.gguf_writer.add_mudd_feed_forward_length(hp["mudd_inter"])
+            self.gguf_writer.add_mudd_tap_count(n_taps)
+            self.gguf_writer.add_mudd_tap_indices(tap_indices)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # tiny mixing weights, kept at full precision as in the reference
+        # the mudd tensors come from a global source tensor, so take the block id from the new name
+        if new_name.startswith("blk."):
+            bid = int(new_name.split(".")[1])
+        for key in (gguf.MODEL_TENSOR.ATTN_GATE, gguf.MODEL_TENSOR.ATTN_VE_GATE, gguf.MODEL_TENSOR.MUDD_DOWN, gguf.MODEL_TENSOR.MUDD_UP):
+            if self.match_model_tensor_name(new_name, key, bid):
+                return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def _swap_rope_pairs(self, w: Tensor, n_head: int) -> Tensor:
+        # HF rotates each pair by -theta; swapping the pair lanes gives the ggml +theta rotation.
+        # q and k are both swapped, so the attention scores do not change.
+        idx = torch.arange(self.head_dim)
+        idx[:2 * self.n_pairs] = idx[:2 * self.n_pairs].view(-1, 2).flip(-1).reshape(-1)
+        w = w.reshape(n_head, self.head_dim, -1)
+        return w[:, idx, :].reshape(n_head * self.head_dim, -1)
+
+    def _pad_cols(self, w: Tensor) -> Tensor:
+        # gates read only the first channels of the input, pad them to n_embd
+        return torch.nn.functional.pad(w, (0, self.hparams["hidden_size"] - w.shape[-1]))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        hp = self.hparams
+        T = gguf.MODEL_TENSOR
+
+        def load(n: str) -> Tensor:
+            return self.model_tensors[n]().float()
+
+        if name == "model.value_embeds.weight":
+            ve = data_torch.reshape(data_torch.shape[0], -1, hp["ve_dim"])[:, :self.n_head_kv]
+            ve = torch.nn.functional.pad(ve, (0, self.head_dim - hp["ve_dim"]))
+            yield self.format_tensor_name(T.VALUE_EMBD), ve.reshape(ve.shape[0], -1) * hp["ve_gate_scale"]
+            return
+
+        if name == "model.mudd.dense1":
+            for il in self.mudd_taps:
+                yield self.format_tensor_name(T.MUDD_DOWN, il), data_torch
+            return
+        if name == "model.mudd.dense2":
+            # rows [0, n) mix the attention input, rows [n, 2n) mix the residual
+            dense2_mlp = load("model.mudd.dense2_mlp")
+            bias       = load("model.mudd.bias")
+            bias_mlp   = load("model.mudd.bias_mlp")
+            for il, taps in self.mudd_taps.items():
+                n = len(taps)
+                w = torch.cat([data_torch[il, :n], dense2_mlp[il, :n]], 0)
+                b = torch.cat([bias[il, :n], bias_mlp[il, :n]], 0)
+                yield self.format_tensor_name(T.MUDD_UP, il), w
+                yield self.format_tensor_name(T.MUDD_UP, il, ".bias"), b
+            return
+        if name.startswith("model.mudd."):
+            return
+
+        if bid is None:
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
+
+        prefix = f"model.layers.{bid}."
+        name = name.removeprefix(prefix)
+
+        if name in ("self_attn.q_proj.weight", "self_attn.k_proj.weight", "self_attn.v_proj.weight"):
+            data_torch = data_torch * load(prefix + "self_attn.qkv_scale")
+            if name == "self_attn.q_proj.weight":
+                data_torch = self._swap_rope_pairs(data_torch, self.n_head)
+            elif name == "self_attn.k_proj.weight":
+                data_torch = self._swap_rope_pairs(data_torch, self.n_head_kv)
+        elif name == "self_attn.o_proj.weight":
+            data_torch = data_torch * load(prefix + "self_attn.o_scale") * load(prefix + "post_lambda_attn") * hp["attn_gate_scale"]
+        elif name == "mlp.down_proj.weight":
+            data_torch = data_torch * load(prefix + "post_lambda_mlp")
+        elif name == "self_attn.attn_gate":
+            yield self.format_tensor_name(T.ATTN_GATE, bid), self._pad_cols(data_torch)
+            return
+        elif name == "self_attn.ve_gate":
+            yield self.format_tensor_name(T.ATTN_VE_GATE, bid), self._pad_cols(data_torch[:self.n_head_kv])
+            return
+        elif name == "self_attn.xsa_alpha":
+            if hp.get("xsa") and bid in hp["xsa_layers"]:
+                yield self.format_tensor_name(T.ATTN_XSA_ALPHA, bid), torch.tanh(data_torch)
+            return
+        elif name == "resid_lambda_attn":
+            yield self.format_tensor_name(T.ATTN_RESID_SCALE, bid), data_torch.reshape(1)
+            return
+        elif name == "resid_lambda_mlp":
+            yield self.format_tensor_name(T.FFN_RESID_SCALE, bid), data_torch.reshape(1)
+            return
+        elif name in ("self_attn.qkv_scale", "self_attn.o_scale", "post_lambda_attn", "post_lambda_mlp"):
+            # folded into the projections above
+            return
+
+        yield from super().modify_tensors(data_torch, prefix + name, bid)
@@ -0,0 +1,204 @@
+from __future__ import annotations
+
+import math
+from typing import Callable, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import LazyTorchTensor, ModelBase, TextModel, gguf
+
+
+@ModelBase.register("Metis16ForCausalLM")
+class Metis16Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.METIS16
+
+    # read by generate_extra_tensors() or not used at inference: dropped in modify_tensors()
+    _folded = {
+        "pass_embeddings", "lm_head.weight", "depth_budget.multiplier",
+        "depth_memory.pass_embeddings", "depth_memory.anchor_embeddings",
+        "depth_memory.state_write.bias", "depth_memory.metadata_write.bias",
+    }
+    _folded_layer = {
+        "mixer_connection.controller.weight", "mixer_connection.controller.bias", "mixer_connection.mix_logits",
+        "mixer_connection.read_logits", "mixer_connection.write_logits",
+        "moe_connection.controller.weight", "moe_connection.controller.bias", "moe_connection.mix_logits",
+        "moe_connection.read_logits", "moe_connection.write_logits",
+        "moe.k_router.weight", "moe.k_router.bias", "moe.k_budget.multiplier",
+    }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hp = self.hparams
+        self.n_pass = hp["num_passes"]
+        self.attn_layers = list(hp["attention_layer_indices"])
+        self._generated: set[str] = set()
+        # one cache slot per pass and layer, while the tensors exist once per physical layer
+        self.block_count = self.n_pass * hp["num_hidden_layers"]
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+        # turns end on the first eos id, so llama.cpp must stop on it too
+        eos = self.hparams.get("eos_token_id")
+        if isinstance(eos, list) and eos:
+            self.gguf_writer.add_eot_token_id(eos[0])
+
+    def set_gguf_parameters(self):
+        hp = self.hparams
+        super().set_gguf_parameters()
+
+        n_layer = hp["num_hidden_layers"]
+        self.gguf_writer.add_head_count_kv([hp["num_key_value_heads"] if il % n_layer in self.attn_layers else 0 for il in range(self.block_count)])
+
+        self.gguf_writer.add_expert_feed_forward_length(hp["expert_intermediate_size"])
+        self.gguf_writer.add_expert_shared_feed_forward_length(hp["shared_expert_intermediate_size"])
+        self.gguf_writer.add_expert_shared_count(1)
+        self.gguf_writer.add_moe_latent_size(hp["latent_dim"])
+
+        d_inner = hp["mamba_expand"] * hp["hidden_size"]
+        self.gguf_writer.add_ssm_conv_kernel(hp["mamba_d_conv"])
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(hp["mamba_d_state"])
+        self.gguf_writer.add_ssm_time_step_rank(d_inner // hp["mamba_head_dim"])
+        self.gguf_writer.add_ssm_group_count(hp["mamba_n_groups"])
+        self.gguf_writer.add_ssm_norm_epsilon(hp["mamba_norm_eps"])
+
+        self.gguf_writer.add_hyper_connection_count(hp["n_streams"])
+        self.gguf_writer.add_metis_pass_count(self.n_pass)
+        self.gguf_writer.add_metis_memory_length(hp["memory_dim"])
+        self.gguf_writer.add_metis_route_length(hp["route_feature_dim"])
+        self.gguf_writer.add_metis_continuation_logit_limit(hp["continuation_logit_limit"])
+        self.gguf_writer.add_attention_pass_lora_rank(hp["attention_pass_lora_rank"])
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        # the checkpoint leaves these names suffix-less; add the suffix of the GGUF tensor so the mapper keeps it
+        if name.endswith((".mixer.impl.mixer.dt_bias", ".moe.selection_bias")):
+            name += ".bias"
+        elif (name.endswith((".moe.expert_embeddings", ".mixer.pass_gates"))
+              or name in ("depth_memory.stream_gate", "stream_embeddings")):
+            name += ".weight"
+        elif name == "depth_memory.stream_gate_bias":
+            name = "depth_memory.stream_gate.bias"
+        return super().filter_tensors((name, gen))
+
+    def _load(self, name: str) -> Tensor:
+        return LazyTorchTensor.to_eager(self.model_tensors[name]())
+
+    def _emit(self, name: str, data: Tensor) -> tuple[str, Tensor]:
+        self._generated.add(name)
+        return name, data
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        hp = self.hparams
+        r = hp["route_feature_dim"]
+        n_stream = hp["n_streams"]
+        T = gguf.MODEL_TENSOR
+        passes = self._load("pass_embeddings")[: self.n_pass]
+
+        # the mixes depend only on the weights and the pass, so Sinkhorn runs once here
+        # the controller runs in the checkpoint dtype, as the reference does
+        for il in range(hp["num_hidden_layers"]):
+            for src, (mix, read, write) in (
+                ("mixer_connection", (T.HC_ATTN_MIX, T.HC_ATTN_READ, T.HC_ATTN_WRITE)),
+                ("moe_connection", (T.HC_FFN_MIX, T.HC_FFN_READ, T.HC_FFN_WRITE)),
+            ):
+                prefix = f"layers.{il}.{src}."
+                ctrl_w, ctrl_b = self._load(prefix + "controller.weight"), self._load(prefix + "controller.bias")
+                mixes, reads, writes = [], [], []
+                for p in range(self.n_pass):
+                    delta = torch.nn.functional.linear(passes[p].to(ctrl_w.dtype), ctrl_w, ctrl_b).float()
+                    m_delta, r_delta, w_delta = torch.split(delta, [n_stream * n_stream, n_stream, n_stream])
+                    log_m = self._load(prefix + "mix_logits").float() + m_delta.view(n_stream, n_stream)
+                    for _ in range(hp["mhc_sinkhorn_iterations"]):
+                        log_m = log_m - torch.logsumexp(log_m, dim=-1, keepdim=True)
+                        log_m = log_m - torch.logsumexp(log_m, dim=-2, keepdim=True)
+                    mixes.append(log_m.exp())
+                    reads.append(torch.softmax(self._load(prefix + "read_logits").float() + r_delta, dim=-1))
+                    writes.append(torch.softmax(self._load(prefix + "write_logits").float() + w_delta, dim=-1))
+                act = ctrl_w.dtype
+                yield self._emit(self.format_tensor_name(mix, il), torch.stack(mixes).to(act).float())
+                yield self._emit(self.format_tensor_name(read, il), torch.stack(reads).to(act).float())
+                yield self._emit(self.format_tensor_name(write, il), torch.stack(writes).to(act).float())
+
+        # one bias per memory entry: write biases, pass and anchor embeddings, and the constant routed width
+        meta_w = self._load("depth_memory.metadata_write.weight").float()
+        mean_k = hp["num_experts_per_tok"] / hp["max_routed_k"]
+        base = (self._load("depth_memory.state_write.bias").float() + self._load("depth_memory.metadata_write.bias").float()
+                + mean_k * meta_w[:, r])
+        pass_e = self._load("depth_memory.pass_embeddings").float()
+        anchor_e = self._load("depth_memory.anchor_embeddings").float()
+        n_anchor = len(self.attn_layers) + 1
+        entry_bias = torch.stack([base + pass_e[p] + anchor_e[a] for p in range(self.n_pass) for a in range(n_anchor)])
+        yield self._emit(self.format_tensor_name(T.DEPTH_MEM_ENTRY_BIAS), entry_bias)
+
+    def _split_features(self, data_torch: Tensor, template: str, parts: tuple[str, ...], d: int, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # the graph reads the concatenated features one block at a time; each block is its own tensor
+        for i, part in enumerate(parts):
+            piece = data_torch[:, i * d:(i + 1) * d] if i < 3 else data_torch[:, 3 * d:]
+            yield from super().modify_tensors(piece, template.format(part), bid)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name in self._generated:
+            yield name, data_torch
+            return
+        if name in self._folded:
+            return
+        hp = self.hparams
+        d = hp["hidden_size"]
+
+        if bid is None:
+            if name == "stream_embeddings.weight":
+                # the graph keeps the streams innermost: [n_stream, n_embd, n_tokens]
+                yield from super().modify_tensors(data_torch.T.contiguous(), name, bid)
+                return
+            if name == "depth_memory.metadata_write.weight":
+                # metadata input becomes [summary, entropy, confidence, continuation]; entropy arrives unnormalized
+                r = hp["route_feature_dim"]
+                meta_w = data_torch.float()
+                entropy_col = meta_w[:, r + 1:r + 2] / math.log(hp["n_routed_experts"])
+                data_torch = torch.cat((meta_w[:, :r], entropy_col, meta_w[:, r + 2:]), dim=1)
+            elif name == "depth_memory.stream_gate.weight":
+                data_torch = data_torch.float() / math.sqrt(d)
+            elif name == "depth_memory.route_projection.weight":
+                yield from self._split_features(data_torch, "depth_memory.route_{}.weight", ("state", "retr", "diff", "hist"), d, bid)
+                return
+            elif name == "continuation.hidden.weight":
+                yield from self._split_features(data_torch, "continuation.hidden_{}.weight", ("state", "retr", "diff", "route"), d, bid)
+                return
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
+
+        local = name.removeprefix(f"layers.{bid}.")
+        if local in self._folded_layer:
+            return
+
+        if local == "mixer.impl.mixer.A_log":
+            data_torch = -torch.exp(data_torch).unsqueeze(-1)
+        elif local == "mixer.impl.mixer.D":
+            data_torch = data_torch.unsqueeze(-1)
+        elif local == "mixer.impl.mixer.conv1d.weight":
+            data_torch = data_torch.squeeze(1)
+        elif local == "mixer.impl.mixer.norm.weight":
+            data_torch = data_torch.reshape(hp["mamba_n_groups"], -1)
+        elif local == "mixer.pass_gates.weight":
+            # the reference applies tanh(gate) in the checkpoint dtype
+            data_torch = torch.tanh(data_torch[: self.n_pass]).to(torch.bfloat16).float()
+        elif local == "moe.shared_expert.gate_up.weight":
+            gate, up = data_torch.chunk(2, dim=0)
+            yield from super().modify_tensors(gate, name.replace("gate_up", "gate"), bid)
+            yield from super().modify_tensors(up, name.replace("gate_up", "up"), bid)
+            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:
+        # FP32 in the checkpoint, tiny and read as rows, or computed here at full precision: keep FP32
+        if (new_name.startswith(("cont_", "depth_mem_entry_bias", "depth_mem_stream_gate"))
+                or new_name.endswith(("ffn_exp_embd.weight", "ffn_gate_inp.weight", "hc_attn_mix.weight", "hc_attn_read.weight",
+                                      "hc_attn_write.weight", "hc_ffn_mix.weight", "hc_ffn_read.weight", "hc_ffn_write.weight"))):
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
@@ -0,0 +1,217 @@
+from __future__ import annotations
+
+from typing import Callable, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, gguf
+
+from .mistral import MistralModel, MistralMoeModel
+
+
+@ModelBase.register("MistralLarge4ForConditionalGeneration")
+class MistralLarge4Model(MistralMoeModel):
+    model_arch = gguf.MODEL_ARCH.MISTRAL_LARGE4
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        n_layer = self.hparams["num_hidden_layers"]
+        dsa = self.hparams["dsa"]
+
+        def expand(pattern: str) -> list[str]:
+            values = pattern.split(",")
+            return [values[i % len(values)] for i in range(n_layer)]
+
+        self.layer_has_rope = [v == "rope" for v in expand(self.hparams.get("nope") or "rope")]
+        self.layer_indexer_full = [v == "full" for v in expand(dsa.get("sharing") or "full")]
+
+        kv_sharing = self.hparams.get("kv_sharing") or {}
+        if "source_layers_str" in kv_sharing:
+            self.layer_kv_source = [int(v) for v in kv_sharing["source_layers_str"].split(",")]
+        else:
+            assert not kv_sharing, f"unsupported kv_sharing config: {kv_sharing}"
+            self.layer_kv_source = list(range(n_layer))
+        assert len(self.layer_kv_source) == n_layer
+
+        # note: the rope dim of the indexer must match the main attention
+        assert dsa.get("rope_head_dim", -1) in (-1, self.hparams["qk_rope_head_dim"])
+        assert dsa.get("match_nope", False), "indexer must follow the NoPE pattern of the main attention"
+        assert dsa.get("ssmax_scaling", False), "indexer without SSMax is not supported"
+        assert self.hparams.get("q_lora_rank") is None
+        assert not self.hparams["yarn"].get("apply_scale", True), "only YaRN without magnitude scaling is supported"
+
+    def dequant_model(self):
+        quant_config = self.hparams.get("quantization")
+        if quant_config is not None:
+            assert quant_config["qformat_weight"] == "fp8_e4m3"
+            assert quant_config.get("qscheme_weight") == "BLOCK"
+            self.hparams["quantization_config"] = {
+                "quant_method": "fp8",
+                "weight_block_size": [quant_config["block_size_out"], quant_config["block_size_in"]],
+            }
+        return super().dequant_model()
+
+    def set_gguf_parameters(self):
+        # skip the MistralMoeModel overrides, they apply the deepseek2 YaRN scaling
+        super(MistralMoeModel, self).set_gguf_parameters()
+        MistralModel.set_mistral_config(self.gguf_writer, self.hparams)
+
+        dsa = self.hparams["dsa"]
+        self.gguf_writer.add_indexer_head_count(dsa["n_heads"])
+        self.gguf_writer.add_indexer_key_length(dsa["head_dim"])
+        self.gguf_writer.add_indexer_top_k(dsa["topk"])
+        self.gguf_writer.add_indexer_types(self.layer_indexer_full)
+        self.gguf_writer.add_layer_norm_eps(dsa.get("norm_eps", 1e-6))
+
+        self.gguf_writer.add_rope_pattern(self.layer_has_rope)
+        self.gguf_writer.add_kv_source_map(self.layer_kv_source)
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        self.gguf_writer.add_final_logit_softcapping(self.hparams["softmax_tanh_gating"])
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        # exported separately with --dspark
+        if name.startswith("mtp_"):
+            return None
+
+        # experts need HF names for the merging logic in DeepseekV2Model
+        # note: bare "layers.N.experts" comes from the NVFP4 quantized_layers config
+        if ".experts." in name or name.endswith(".experts"):
+            name = "model." + name.replace(".experts", ".mlp.experts")
+            name = name.replace(".w1.", ".gate_proj.").replace(".w2.", ".down_proj.").replace(".w3.", ".up_proj.")
+
+        # skip the MistralMoeModel renaming, the qscale names are needed for block FP8 dequant
+        return super(MistralMoeModel, cls).filter_tensors((name, gen))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "lerope.alpha":
+            # LeRoPE scales inv_freq by exp(alpha), ggml divides theta by the freq factors
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.exp(-data_torch.float()))
+            return
+
+        if bid is not None and name.endswith((".ssmax_scaling.s", ".ssmax_scaling.b")):
+            tensor = gguf.MODEL_TENSOR.INDEXER_SSMAX if ".indexer." in name else gguf.MODEL_TENSOR.ATTN_SSMAX
+            suffix = ".weight" if name.endswith(".s") else ".bias"
+            yield (self.format_tensor_name(tensor, bid, suffix), data_torch.float())
+            return
+
+        if bid is not None and ".indexer.q_input_norm." in name:
+            suffix = ".weight" if name.endswith(".weight") else ".bias"
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_NORM, bid, suffix), data_torch)
+            return
+
+        if bid is not None and ".indexer." in name:
+            name = "model." + name.replace(".attention.", ".self_attn.")
+
+        # NoPE layers have a larger non-rotary head, so wkv_b is split per layer
+        if bid is not None and name.endswith(".wkv_b.weight"):
+            n_head = self.hparams["num_attention_heads"]
+            v_head_dim = self.hparams["v_head_dim"]
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+            if not self.layer_has_rope[bid]:
+                qk_nope_head_dim += self.hparams["qk_rope_head_dim"]
+
+            assert data_torch.shape[0] == n_head * (qk_nope_head_dim + v_head_dim)
+
+            kv_b = data_torch.view(n_head, qk_nope_head_dim + v_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K_B, bid), k_b)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V_B, bid), v_b)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+class MistralLarge4DSparkModel(MistralLarge4Model):
+    """Mistral Large 4 DSpark draft (mtp_dspark.*) as a DFLASH GGUF, export with --dspark"""
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    _DSPARK_ROOT: dict[str, gguf.MODEL_TENSOR] = {
+        "backbone.target_projection.weight": gguf.MODEL_TENSOR.FC,
+        "backbone.target_norm.weight":       gguf.MODEL_TENSOR.ENC_OUTPUT_NORM,
+        "backbone.final_norm.weight":        gguf.MODEL_TENSOR.OUTPUT_NORM,
+        "backbone.mask_embedding":           gguf.MODEL_TENSOR.DSPARK_MASK_EMBD,
+        "autoregressive_head.embedding.weight": gguf.MODEL_TENSOR.DSPARK_MARKOV_W1,
+        "autoregressive_head.output.weight":    gguf.MODEL_TENSOR.DSPARK_MARKOV_W2,
+        "autoregressive_head.joint_proj.weight": gguf.MODEL_TENSOR.DSPARK_JOINT_PROJ,
+        "autoregressive_head.joint_proj.bias":   gguf.MODEL_TENSOR.DSPARK_JOINT_PROJ,
+    }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        mtp = self.hparams.get("mtp") or {}
+        dspark = mtp.get("dspark")
+        if dspark is None:
+            raise ValueError("Mistral Large 4 DSpark export needs the mtp.dspark config")
+        assert dspark.get("head_type") == "rnn", f"unsupported DSpark head type: {dspark.get('head_type')}"
+        assert dspark.get("draft_input") == "position_specific", f"unsupported DSpark draft input: {dspark.get('draft_input')}"
+
+        windows = {int(v) for v in (mtp.get("ragged_attention") or "").split(",") if v.strip()}
+        assert len(windows) == 1, f"DSpark draft layers must share one sliding window, got {mtp.get('ragged_attention')}"
+
+        self._n_stage = int(mtp["n_layers"])
+        self._depth = int(mtp["depth"])
+        self._window = windows.pop()
+        self._target_layers = [int(v) for v in mtp["target_activation_layers"].split(",")]
+
+        self.layer_has_rope = [True] * self._n_stage
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if not name.startswith("mtp_dspark."):
+            return None
+        name = name.removeprefix("mtp_dspark.")
+        if name.startswith("backbone.layers."):
+            return name.removeprefix("backbone."), gen
+        return f"dspark.{name}", gen
+
+    def set_vocab(self):
+        super().set_vocab()
+        # block uses mask_embd, mask token id is not used
+        self.gguf_writer.add_mask_token_id(0)
+
+    def set_gguf_parameters(self):
+        hp = self.hparams
+        w = self.gguf_writer
+        w.add_block_count(self._n_stage)
+        w.add_context_length(hp["max_position_embeddings"])
+        w.add_embedding_length(hp["hidden_size"])
+        w.add_feed_forward_length(hp["intermediate_size"])
+        w.add_head_count(hp["num_attention_heads"])
+        w.add_head_count_kv(1)
+        w.add_key_length(hp["kv_lora_rank"] + hp["qk_rope_head_dim"])
+        w.add_value_length(hp["kv_lora_rank"])
+        w.add_key_length_mla(hp["qk_nope_head_dim"] + hp["qk_rope_head_dim"])
+        w.add_value_length_mla(hp["v_head_dim"])
+        w.add_kv_lora_rank(hp["kv_lora_rank"])
+        w.add_rope_dimension_count(hp["qk_rope_head_dim"])
+        w.add_rope_freq_base(float(hp["rope_theta"]))
+        w.add_layer_norm_rms_eps(hp["rms_norm_eps"])
+        MistralModel.set_mistral_config(w, hp)
+        w.add_sliding_window(self._window)
+        w.add_final_logit_softcapping(hp["softmax_tanh_gating"])
+        w.add_block_size(self._depth)
+        # draft reads layer output = input of next layer
+        w.add_target_layers([layer + 1 for layer in self._target_layers])
+        w.add_sample_from_anchor(True)
+        w.add_has_confidence_head(False)
+        w.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("dspark."):
+            key = name.removeprefix("dspark.")
+            if key not in self._DSPARK_ROOT:
+                raise ValueError(f"Unexpected DSpark tensor {name!r}")
+            suffix = ".bias" if key.endswith(".bias") else ".weight"
+            yield (self.format_tensor_name(self._DSPARK_ROOT[key], None, suffix), data_torch)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -428,11 +428,19 @@ class NemotronHModel(GraniteHybridModel):
    is_moe: bool = False
    supports_mtp_export = True
    _experts: list[dict[str, Tensor]] | None = None
+    # tokens past the embedding rows (left out of the vocab), and special tokens that are printed as text
+    _vocab_skip: tuple[str, ...] = ()
+    _user_defined_tokens: re.Pattern[str] | None = None

    _SSM_LAYER_TYPES = {"mamba", "linear_attention"}
    _ATTN_LAYER_TYPES = {"attention", "full_attention"}
    _MLP_LAYER_TYPES = {"moe"}

+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        if ".mixer.experts." in name:
+            name = name.replace("backbone.", "model.").replace(".mixer.experts.", ".mlp.experts.")
+        return super()._map_fp8_weight_names(name)
+
    def __init__(self, *args, **kwargs):
        # We have to determine the correct model architecture (MoE vs non-MoE) before
        # calling the parent __init__. This is because the parent constructor
@@ -611,7 +619,11 @@ class NemotronHModel(GraniteHybridModel):
        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
        self.hparams["vocab_size"] = vocab_size

-        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
+        # draft-model converters (DFlash, LiLiCorr) call this with their own instance, which has neither hook
+        vocab_skip = getattr(self, "_vocab_skip", ())
+        user_defined_tokens = getattr(self, "_user_defined_tokens", None)
+
+        assert max(id_ for tok, id_ in tokenizer.vocab.items() if tok not in vocab_skip) < vocab_size  # ty: ignore[unresolved-attribute]

        tokpre = self.get_vocab_base_pre(tokenizer)

@@ -633,7 +645,8 @@ class NemotronHModel(GraniteHybridModel):
                        if previous_token != token:
                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

-                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                    printed = user_defined_tokens is not None and user_defined_tokens.fullmatch(token)
+                    if (added_tokens_decoder[i].special or self.does_token_look_special(token)) and not printed:
                        toktypes.append(gguf.TokenType.CONTROL)
                    else:
                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
@@ -829,3 +842,140 @@ class NemotronHPuzzleModel(NemotronHModel):
        if item[0].startswith("mtp."):
            return None
        return super().filter_tensors(item)
+
+
+# NeMo speechlm2 SALM checkpoint: a Nemotron-H decoder under llm.*, the speech encoder under perception.*
+@ModelBase.register("Nemotron3_5TranscribeForConditionalGeneration")
+@ModelBase.example("nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B")
+class NemotronTranscribeModel(NemotronHModel):
+    model_arch = gguf.MODEL_ARCH.NEMOTRON_H
+    # the audio placeholder (id vocab_size) has no embedding row; speaker, event and style tags print with the transcript
+    _vocab_skip = ("<|audio|>",)
+    _user_defined_tokens = re.compile(r"<spk:\d+>|<ev:[a-z_]+>|<sty:(?:stt|end):[a-z_]+>")
+    _renames = {
+        "llm.model.embed_tokens.weight": "backbone.embeddings.weight",
+        "llm.model.norm.weight":         "backbone.norm_f.weight",
+        "llm.lm_head.weight":            "lm_head.weight",
+    }
+    # grouped experts: [n_expert, n_in, n_out] -> [n_expert, n_out, n_in]
+    _grouped_experts = {"gate_and_up_projs": "up_proj", "down_projs": "down_proj"}
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if name.startswith(("perception.", "ctc_timestamp.")) or "_extra_state" in name:
+            return None
+        if not name.startswith("llm."):
+            # the FP8 and NVFP4 releases store the decoder as backbone.*
+            raise ValueError(f"unexpected tensor {name}: only the BF16 checkpoint layout (llm.*) is supported")
+        if name in cls._renames:
+            name = cls._renames[name]
+        elif name.startswith("llm.model.layers."):
+            name = "backbone.layers." + name[len("llm.model.layers."):]
+        elif name.startswith("llm.mtp.layers."):
+            name = "mtp.layers." + name[len("llm.mtp.layers."):]
+        else:
+            raise ValueError(f"unexpected decoder tensor {name}")
+        return super().filter_tensors((name.replace("._fp32_params.", "."), gen))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        leaf = name.rsplit(".", 1)[-1]
+        if ".mixer.experts." in name and leaf in self._grouped_experts:
+            if name.startswith("mtp."):
+                assert self._mtp_bid is not None, f"{name}: MTP head is not exported"
+                bid = self._mtp_bid
+            assert bid is not None
+            data_torch = data_torch.permute(0, 2, 1).contiguous()
+            name = f"model.layers.{bid}.mlp.experts.{self._grouped_experts[leaf]}.weight"
+            yield from ModelBase.modify_tensors(self, data_torch, name, bid)
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Nemotron3_5TranscribeForConditionalGeneration")
+class NemotronTranscribeAudioModel(MmprojModel):
+    # NeMo ParallelExpertEncoder: ASR and speaker transformers, speaker activity fused into the ASR states
+    has_vision_encoder = False
+    has_audio_encoder = True
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        asr = self.global_config["pe_encoder_config"]["asr_encoder_cfg"]
+        return {
+            "hidden_size":         asr["d_model"],
+            "intermediate_size":   int(float(asr.get("ff_expansion", 4)) * asr["d_model"]),
+            "num_hidden_layers":   asr["n_layers"],
+            "num_attention_heads": asr["n_heads"],
+        }
+
+    @staticmethod
+    def _rope_dims(cfg: dict[str, Any]) -> int:
+        assert cfg.get("self_attention_model") == "rope" and cfg.get("subsampling", "feature_stacking") == "feature_stacking"
+        assert cfg.get("attn_mode", "full") == "full" and not cfg.get("xscaling", False) and cfg.get("pre_block_norm", True)
+        assert float(cfg.get("rope_base", 10000.0)) == 10000.0
+        n_rot = int(cfg["d_model"] // cfg["n_heads"] * float(cfg.get("rotary_fraction", 1.0)))
+        assert n_rot >= 2 and n_rot % 2 == 0
+        return n_rot
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        pe = self.global_config["pe_encoder_config"]
+        asr = pe["asr_encoder_cfg"]
+        diar = pe["diarization_model_cfg"]
+        spk = diar["encoder"]
+        assert pe.get("asr_encoder_type") == "transformer" and pe.get("speaker_feature_mode", "thresholded") == "thresholded"
+        assert pe.get("asr_normalize_type") == "per_feature" and pe.get("diar_normalize_type") in (None, "NA")
+        assert asr["subsampling_factor"] == spk["subsampling_factor"] == 8 and asr["feat_in"] == spk["feat_in"]
+        assert self._rope_dims(spk) == spk["d_model"] // spk["n_heads"] and not spk.get("qkv_bias", False)
+        assert diar.get("high_resolution", False)
+        # constants in clip: the ParallelExpertEncoder online settings and the Sortformer streaming config
+        assert (pe.get("online_inference_length", 500), pe.get("chunk_left_context", 50), pe.get("chunk_right_context", 50),
+                pe.get("diar_spkcache_len", 188), pe.get("diar_fifo_len", 40), pe.get("diar_spkcache_update_period", 300),
+                pe.get("speaker_activity_threshold", 0.5)) == (500, 50, 50, 188, 40, 300, 0.5), \
+            "streaming settings differ from the constants in tools/mtmd/models/nemotron-transcribe.cpp"
+        sm = diar["sortformer_modules"]
+        assert (sm["spkcache_sil_frames_per_spk"], sm["pred_score_threshold"], sm["scores_boost_latest"], sm["strong_boost_rate"],
+                sm["weak_boost_rate"], sm["min_pos_scores_rate"], sm["use_learnable_sil_emb"]) == (1, 0.25, 0.05, 0.75, 1.5, 0.5, True), \
+            "speaker cache settings differ from the constants in tools/mtmd/models/nemotron-transcribe.cpp"
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_TRANSCRIBE)
+        self.gguf_writer.add_audio_num_mel_bins(asr["feat_in"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+        self.gguf_writer.add_audio_subsampling_factor(asr["subsampling_factor"])
+        self.gguf_writer.add_audio_rope_dims(self._rope_dims(asr))
+        self.gguf_writer.add_audio_spk_block_count(spk["n_layers"])
+        self.gguf_writer.add_audio_spk_head_count(spk["n_heads"])
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # ParallelExpertEncoder._build_sinusoid_position_encoding
+        pe = self.global_config["pe_encoder_config"]
+        n_spk = int(pe["diarization_model_cfg"]["sortformer_modules"]["num_spks"])
+        d = int(pe["asr_encoder_cfg"]["d_model"])
+        scale = float(pe.get("spk_kernel_scale", 1.0))
+        pos = torch.arange(n_spk, dtype=torch.float32)[:, None]
+        div = torch.exp(torch.arange(0, d, 2, dtype=torch.float32) * -(np.log(10000.0) / d))
+        kernel = torch.zeros(n_spk, d, dtype=torch.float32)
+        kernel[:, 0::2] = torch.sin(pos * div)
+        kernel[:, 1::2] = torch.cos(pos * div)
+        # stored as [d, n_spk]: the graph multiplies it with the speaker activity directly
+        yield self.format_tensor_name(gguf.MODEL_TENSOR.A_FUSION_SPK_KERNEL, suffix=""), (kernel * scale).T.contiguous()
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.A_SPK_HEAD_UPSAMPLE, None):
+            return gguf.GGMLQuantizationType.F16  # ggml_conv_1d kernel
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if not name.startswith("perception.") or "_extra_state" in name:
+            return None
+        # training-only head; the speaker model's unused front end
+        if name.startswith(("perception.encoder.diarization_model.sortformer_modules.hidden_to_spks.",
+                            "perception.encoder.diarization_model.preprocessor.")):
+            return None
+        return name, gen
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "perception.preprocessor.featurizer.fb":
+            data_torch = data_torch.squeeze(0)
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -5,6 +5,7 @@ from pathlib import Path
 from typing import Any, Iterable, TYPE_CHECKING

 import torch
+import torch.nn.functional as F

 if TYPE_CHECKING:
    from torch import Tensor
@@ -45,6 +46,102 @@ def _tensor_shapes(dir_model: Path) -> dict[str, tuple[int, ...]]:
        return {name: tuple(part[name].shape) for name in part.keys()}


+def _has_voice_encoder(dir_model: Path) -> bool:
+    # without voice cloning, the mimi encoder weights are all zero
+    part_names = ModelBase.get_model_part_names(dir_model, "model", ".safetensors")
+    with gguf.utility.SafetensorsLocal(dir_model / part_names[0]) as part:
+        nonzero = {bool(part[name].mmap_bytes().any()) for name in part.keys() if name.startswith("mimi.encoder")}
+    if len(nonzero) != 1:
+        raise ValueError(f"{dir_model}: the mimi encoder is partly zeroed out")
+    return nonzero.pop()
+
+
+def _load_safetensors(path: Path, prefixes: tuple[str, ...]) -> dict[str, Tensor]:
+    dtypes = {"F32": torch.float32, "BF16": torch.bfloat16, "I64": torch.int64}
+    with gguf.utility.SafetensorsLocal(path) as part:
+        return {name: torch.from_numpy(part[name].mmap_bytes()).view(dtypes[part[name].dtype]).reshape(part[name].shape)
+                for name in part.keys() if name.startswith(prefixes)}
+
+
+def _rope_pairs(x: Tensor, n_head: int, rope_theta: float) -> Tensor:
+    # ggml NORM mode rope, x: [n_pos, n_embd]
+    n_pos, n_embd = x.shape
+    head_dim = n_embd // n_head
+    angle = torch.arange(n_pos, dtype=x.dtype)[:, None, None] * \
+        torch.pow(rope_theta, -torch.arange(0, head_dim, 2, dtype=x.dtype) / head_dim)
+    x = x.view(n_pos, n_head, head_dim // 2, 2)
+    return torch.stack([x[..., 0] * angle.cos() - x[..., 1] * angle.sin(),
+                        x[..., 0] * angle.sin() + x[..., 1] * angle.cos()], -1).reshape(n_pos, n_embd)
+
+
+def _voice_prompt_embd(weights: dict[str, Tensor], kv: list[Tensor], n_head: int, rope_theta: float) -> tuple[Tensor, float]:
+    # the voice state only has K/V of each layer (kv[il] is [k/v, pos, n_embd]): layer 0 V gives each frame up to its scale, layer 1 V gives the scale
+    # returns the frames without bos_before_voice, and the max relative K/V error when the frames are prompted again
+    w = {name: t.double() for name, t in weights.items()}
+    kv = [t.double() for t in kv]
+    k0, v0, v1 = kv[0][0], kv[0][1], kv[1][1]
+    n_pos, n_embd = v0.shape
+    head_dim = n_embd // n_head
+
+    def p(il: int, name: str) -> Tensor:
+        return w[f"flow_lm.transformer.layers.{il}.{name}"]
+
+    def norm(x: Tensor, il: int, name: str) -> Tensor:
+        return F.layer_norm(x, (n_embd,), p(il, f"{name}.weight"), p(il, f"{name}.bias"), 1e-5)
+
+    def heads(t: Tensor) -> Tensor:
+        return t.reshape(n_pos, n_head, head_dim).transpose(0, 1)
+
+    def rows(t: Tensor) -> Tensor:
+        return t.transpose(0, 1).reshape(n_pos, n_embd)
+
+    def attn(il: int, h: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        q = _rope_pairs(h @ p(il, "self_attn.in_proj.weight").chunk(3)[0].T, n_head, rope_theta)
+        return rows(F.scaled_dot_product_attention(heads(q), heads(k), heads(v), is_causal=True)) @ p(il, "self_attn.out_proj.weight").T
+
+    def ffn(il: int, x: Tensor) -> Tensor:
+        return F.gelu(norm(x, il, "norm2") @ p(il, "linear1.weight").T) @ p(il, "linear2.weight").T
+
+    wv = p(0, "self_attn.in_proj.weight").chunk(3)[2]
+    h = torch.linalg.solve(wv, v0.T).T                   # norm1(x)
+    z = (h - p(0, "norm1.bias")) / p(0, "norm1.weight")  # (x - mean(x)) / std(x)
+
+    # fit z on [speaker_proj, 1] to get the latent direction back
+    proj = w["flow_lm.speaker_proj_weight"]
+    fit = torch.linalg.lstsq(torch.cat([proj, torch.ones(n_embd, 1, dtype=proj.dtype)], 1), z[1:].T).solution
+    frames = (proj @ fit[:-1]).T
+
+    a0 = attn(0, h, k0, v0)[1:]
+    wv1 = p(1, "self_attn.in_proj.weight").chunk(3)[2]
+
+    def err(scale: Tensor) -> Tensor:  # per frame relative error of layer 1 V
+        x = scale[:, None] * frames + a0
+        x = x + ffn(0, x)
+        return (norm(x, 1, "norm1") @ wv1.T - v1[1:]).norm(dim=1) / v1[1:].norm(dim=1)
+
+    grid = torch.logspace(-3, 3, 49, dtype=torch.float64)
+    best = torch.stack([err(s.expand(n_pos - 1)) for s in grid]).argmin(0)
+    lo, hi = grid[(best - 1).clamp(min=0)], grid[(best + 1).clamp(max=len(grid) - 1)]
+    for _ in range(40):  # golden section search, all frames at once
+        m1, m2 = hi - 0.618 * (hi - lo), lo + 0.618 * (hi - lo)
+        left = err(m1) < err(m2)
+        lo, hi = torch.where(left, lo, m1), torch.where(left, m2, hi)
+    frames = (lo + hi)[:, None] / 2 * frames
+
+    # prompt the frames again and compare K/V with the shipped state
+    x = torch.cat([w["flow_lm.bos_before_voice"].reshape(1, n_embd), frames])
+    drift = 0.0
+    for il, (k_ref, v_ref) in enumerate(kv):
+        h = norm(x, il, "norm1")
+        _, wk, wv = p(il, "self_attn.in_proj.weight").chunk(3)
+        k, v = _rope_pairs(h @ wk.T, n_head, rope_theta), h @ wv.T
+        for a, b in ((k, k_ref), (v, v_ref)):
+            drift = max(drift, ((a - b).norm(dim=1) / b.norm(dim=1)).median().item())
+        x = x + attn(il, h, k, v)
+        x = x + ffn(il, x)
+    return frames, drift
+
+
@ModelBase.register_hparams_loader(lambda dir_model: "flow_lm.bos_emb" in _tensor_shapes(dir_model))
 def _load_hparams(dir_model: Path) -> dict[str, Any]:
    logger.info("gguf: detected pocket-tts checkpoint, deriving hparams from tensor shapes")
@@ -194,22 +291,27 @@ class PocketTTSMmprojModel(MmprojModel):
        (gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_Q, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_K, gguf.MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_V),
    )

+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.has_audio_encoder = _has_voice_encoder(self.dir_model)
+
    def set_gguf_parameters(self):
        self.gguf_writer.add_file_type(self.ftype)
        assert self.hparams_audio is not None

        # voice-prompt encoder: mimi encoder + speaker_proj
-        self.gguf_writer.add_clip_has_audio_encoder(True)
-        # note: the 24kHz sample rate is hardcoded on the clip.cpp side, like the other audio models
-        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_SPKENC)
-        self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
-        self.gguf_writer.add_audio_block_count(self.hparams_audio["num_hidden_layers"])
-        self.gguf_writer.add_audio_embedding_length(self.hparams_audio["hidden_size"])
-        self.gguf_writer.add_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
-        self.gguf_writer.add_audio_head_count(self.hparams_audio["num_attention_heads"])
-        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
-        # mimi convolves the waveform directly, it is passed around as a 1-row "mel"
-        self.gguf_writer.add_audio_num_mel_bins(1)
+        if self.has_audio_encoder:
+            self.gguf_writer.add_clip_has_audio_encoder(True)
+            # note: the 24kHz sample rate is hardcoded on the clip.cpp side, like the other audio models
+            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.POCKETTTS_SPKENC)
+            self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+            self.gguf_writer.add_audio_block_count(self.hparams_audio["num_hidden_layers"])
+            self.gguf_writer.add_audio_embedding_length(self.hparams_audio["hidden_size"])
+            self.gguf_writer.add_audio_feed_forward_length(self.hparams_audio["intermediate_size"])
+            self.gguf_writer.add_audio_head_count(self.hparams_audio["num_attention_heads"])
+            self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+            # mimi convolves the waveform directly, it is passed around as a 1-row "mel"
+            self.gguf_writer.add_audio_num_mel_bins(1)

        # generation: flow-matching decoder + mimi decoder
        # the SEANet and flow net hparams are constant across the family, clip.cpp holds them
@@ -229,12 +331,57 @@ class PocketTTSMmprojModel(MmprojModel):
        # conv1d/conv1d_dw kernels must be F16, ggml_conv_1d(_dw) has no BF16 path
        if ".seanet." in new_name or new_name in ("a.downsample.conv.weight", "a.gen.wav.upsample.weight"):
            return gguf.GGMLQuantizationType.F16
+        # voices are read on the host
+        if new_name.startswith("voice."):
+            return gguf.GGMLQuantizationType.F32
        return False

+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # embeddings/<name>.safetensors hold the K/V state after the voice prompt (see get_state_for_audio_prompt() in the reference), turn it back into prompt embeddings
+        paths = sorted((self.dir_model / "embeddings").glob("*.safetensors"))
+        if not paths:
+            return
+        weights = _load_safetensors(self.dir_model / "model.safetensors",
+                                    ("flow_lm.transformer.layers.", "flow_lm.speaker_proj_weight", "flow_lm.bos_before_voice"))
+        n_layer = self.global_config["num_hidden_layers"]
+        for path in paths:
+            state = _load_safetensors(path, ("transformer.layers.",))
+            kv = []
+            for il in range(n_layer):
+                prefix = f"transformer.layers.{il}.self_attn/"
+                if any(prefix + k not in state for k in ("cache", "offset")):
+                    break
+                n_pos = int(state[prefix + "offset"].item())
+                # older packs have no pad entry
+                if prefix + "pad" in state and int(state[prefix + "pad"].item()) != 0:
+                    break
+                kv.append(state[prefix + "cache"][:, 0, :n_pos].flatten(2))  # [k/v, pos, n_embd]
+            if len(kv) != n_layer or any(t.shape != kv[0].shape for t in kv) or kv[0].shape[1] < 2 or \
+                    kv[0].shape[2] != self.global_config["hidden_size"]:
+                logger.warning(f"voice {path.stem}: skipped, not an unpadded voice state of this backbone")
+                continue
+            embd, drift = _voice_prompt_embd(weights, kv, self.global_config["num_attention_heads"],
+                                             self.global_config["rope_theta"])
+            # some packs ship a state made with other weights, the reference still uses it: keep up to 50% error, warn above 5%
+            if drift > 0.5:
+                logger.warning(f"voice {path.stem}: skipped, its state does not belong to these weights (error {drift:.1e})")
+                continue
+            log = logger.warning if drift > 0.05 else logger.info
+            log(f"voice {path.stem}: {embd.shape[0]} frames, prompting them again gives the shipped state within {drift:.1e}")
+            yield (f"voice.{path.stem}", embd.float())
+
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        del bid  # the block index of the mimi transformers is parsed here, not by the base class
        T = gguf.MODEL_TENSOR

+        if name.startswith("voice."):
+            yield (name, data_torch)
+            return
+
+        if not self.has_audio_encoder and (name.startswith(("mimi.encoder", "mimi.downsample."))
+                                           or name == "flow_lm.speaker_proj_weight"):
+            return  # voice-prompt encoder
+
        if name in ("flow_lm.bos_emb", "flow_lm.bos_before_voice", "flow_lm.conditioner.embed.weight"):
            return  # folded into the backbone embedding table
        if name.startswith("flow_lm.transformer.") or name.startswith("flow_lm.out_norm."):
@@ -83,6 +83,15 @@ class Qwen2Model(TextModel):
 class Qwen2MoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.QWEN2MOE

+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        if name.removesuffix(".weight").endswith(".mlp.experts.gate_up_proj"):
+            bid = next(int(part) for part in self.map_tensor_name(name).split(".") if part.isdecimal())
+            return (
+                self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid),
+                self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid),
+            )
+        return super()._map_fp8_weight_names(name)
+
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
@@ -382,6 +391,16 @@ class _QwenMtpMixin:
 class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
    model_arch = gguf.MODEL_ARCH.QWEN3NEXT

+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        if name.endswith(".linear_attn.in_proj_qkvz.weight"):
+            bid = next(int(part) for part in self.map_tensor_name(name).split(".") if part.isdecimal())
+            # Both split projections keep the original per-tensor scales.
+            return (
+                self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, bid),
+                self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_GATE, bid),
+            )
+        return super()._map_fp8_weight_names(name)
+
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        self.gguf_writer.add_ssm_conv_kernel(self.hparams["linear_conv_kernel_dim"])
@@ -0,0 +1,92 @@
+from __future__ import annotations
+
+import math
+
+from typing import Iterable, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, TextModel, gguf, logger
+
+
+@ModelBase.register("RicursiveForCausalLM")
+@ModelBase.example("ricursiveAI/Ripple-65B-Preview")
+class RicursiveModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.RICURSIVE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hparams = self.hparams
+
+        if (act := hparams.get("hidden_act", "silu")) != "silu":
+            raise ValueError(f"Ricursive: unsupported hidden_act {act!r}")
+
+        # RicursiveConfig derives the layer pattern from layer_types alone, by default full attention iff i % 5 == 4
+        if (layer_types := hparams.get("layer_types")) is None:
+            layer_types = ["full_attention" if i % 5 == 4 else "sliding_attention" for i in range(self.block_count)]
+        if len(layer_types) != self.block_count:
+            raise ValueError(f"Ricursive: layer_types has {len(layer_types)} entries, expected {self.block_count}")
+        if unknown := set(layer_types) - {"sliding_attention", "full_attention"}:
+            raise ValueError(f"Ricursive: unsupported layer_types {sorted(unknown)}")
+        self.is_swa = [t == "sliding_attention" for t in layer_types]
+        # only sliding layers use RoPE, an all-full-attention model would have no positional encoding
+        if not any(self.is_swa):
+            raise ValueError("Ricursive: layer_types has no sliding_attention layer, all-full-attention models are not supported")
+        # RicursiveConfig pops a stored hybrid_layer_pattern and recomputes it from layer_types
+        if (pattern := hparams.get("hybrid_layer_pattern")) is not None and list(pattern) != [int(swa) for swa in self.is_swa]:
+            logger.warning("Ricursive: hybrid_layer_pattern does not match layer_types, ignoring it")
+
+        if not hparams.get("gate_full_precision", True):
+            logger.warning("Ricursive: gate_full_precision is false, llama.cpp computes the attention gate in f32")
+
+        # RicursiveConfig defaults; the reference ignores the tensors of a disabled feature
+        self.use_valres = hparams.get("use_valres", False)
+        self.use_gate = hparams.get("use_gated_attention", True)
+        names = list(self.model_tensors)
+        if bias := [name for name in names if name.endswith("bias")]:
+            raise ValueError(f"Ricursive: bias tensors are not supported, found {bias[0]!r}")
+        for enabled, suffix in ((self.use_valres, ".self_attn.value_res_lambda"), (self.use_gate, ".self_attn.g_proj.weight")):
+            if enabled and sum(name.endswith(suffix) for name in names) != self.block_count:
+                raise ValueError(f"Ricursive: expected one {suffix!r} tensor per layer")
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+        # eos_token_id is [<|im_end|>, <|endoftext|>]: register the turn end as EOT so generation stops on it
+        eos = self.hparams.get("eos_token_id")
+        if isinstance(eos, list):
+            extra = [e for e in eos if e != self.hparams.get("bos_token_id")]
+            if extra:
+                self.gguf_writer.add_eot_token_id(extra[0])
+                logger.info(f"gguf: eot_token_id = {extra[0]}")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["head_dim"])
+
+        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern(self.is_swa)
+        # sliding layers use RoPE, full attention layers use no positional encoding
+        self.gguf_writer.add_rope_pattern(self.is_swa)
+
+        if (embed_scale := hparams.get("embed_scale")) is not None:
+            if embed_scale == 0:
+                raise ValueError(f"Ricursive: unsupported embed_scale {embed_scale!r}")
+            self.gguf_writer.add_embedding_scale(float(embed_scale))
+        # absent and 0 both mean no clamp (the reference tests `if clamp_threshold:`)
+        if (clamp := hparams.get("ffn_clamp_threshold")) is not None:
+            if clamp < 0 or not math.isfinite(clamp):
+                raise ValueError(f"Ricursive: unsupported ffn_clamp_threshold {clamp!r}")
+            self.gguf_writer.add_swiglu_clamp([float(clamp)] * self.block_count)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # value_res_lambda has no .weight suffix in the checkpoint
+        if name.endswith(".value_res_lambda"):
+            if not self.use_valres:
+                return
+            name += ".weight"
+        elif name.endswith(".g_proj.weight") and not self.use_gate:
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -0,0 +1,133 @@
+from __future__ import annotations
+
+import re
+from typing import Callable, Iterable, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, gguf
+from .deepseek import DeepseekV2Model
+
+
+@ModelBase.register("Xing4_0ForCausalLM")
+class Xing4_0Model(DeepseekV2Model):
+    """Xing4_0 : DeepSeek-V2/V3 backbone (MLA + MoE) + mHC residual mixing."""
+
+    model_arch = gguf.MODEL_ARCH.XING4_0
+
+    skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
+
+    # map (prefix, kind) -> MODEL_TENSOR enum
+    _hc_tensor_map = {
+        ("hc_attn", "fn"):    gguf.MODEL_TENSOR.HC_ATTN_FN,
+        ("hc_attn", "base"):  gguf.MODEL_TENSOR.HC_ATTN_BASE,
+        ("hc_attn", "scale"): gguf.MODEL_TENSOR.HC_ATTN_SCALE,
+        ("hc_ffn",  "fn"):    gguf.MODEL_TENSOR.HC_FFN_FN,
+        ("hc_ffn",  "base"):  gguf.MODEL_TENSOR.HC_FFN_BASE,
+        ("hc_ffn",  "scale"): gguf.MODEL_TENSOR.HC_FFN_SCALE,
+    }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        # needed by filter_tensors() before any tensor is filtered
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        # the NextN/MTP block lives past num_hidden_layers (model.layers.40 -> blk.40)
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        # --no-mtp: drop the appended NextN block entirely.
+        if is_mtp and cls.no_mtp:
+            return None
+        # --mtp: keep ONLY NextN-block tensors plus the shared embeddings/
+        # norm/lm_head (so the resulting GGUF carries just the draft head).
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
+    def set_gguf_parameters(self):
+        rope_type = self.rope_parameters.get("rope_type") or self.rope_parameters.get("type")
+        if rope_type == "rope":
+            self.rope_parameters["rope_type"] = "yarn"
+            if "type" in self.rope_parameters:
+                self.rope_parameters["type"] = "yarn"
+            # base.py's YARN branch accesses this with a direct key lookup
+            self.rope_parameters.setdefault("original_max_position_embeddings",
+                                            self.hparams.get("original_max_position_embeddings", 4096))
+
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        # NextN/MTP prediction layers
+        if not self.no_mtp and (num_nextn_predict_layers := hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+        self.gguf_writer.add_hyper_connection_count(
+            hparams.get("hc_mult",  1))
+        self.gguf_writer.add_hyper_connection_sinkhorn_iterations(
+            hparams.get("hc_sinkhorn_iters",  20))
+        self.gguf_writer.add_hyper_connection_epsilon(
+            hparams.get("hc_eps",  1e-6))
+
+    def set_vocab(self):
+        # Xing4_0 uses a SentencePiece tokenizer (tokenizer.model + Xing4_0Tokenizer).
+        # V2's set_vocab tries GPT2/BPE first, which fails for SPM. Use the SPM path directly.
+        # Note: no add_tokenizer_pre override — the C++ SPM load path ignores
+        # tokenizer.ggml.pre entirely (pre-tokenizers only apply to BPE vocabs).
+        self._set_vocab_sentencepiece()
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # handle mHC tensors: model.layers.{N}.{attn_hc|ffn_hc}.{hc_fn|hc_base|hc_scale}
+        match = re.match(r"model\.layers\.(\d+)\.(attn_hc|ffn_hc)\.(.+)$", name)
+        if match:
+            layer_idx = int(match.group(1))
+            hc_type = match.group(2)  # "attn_hc" or "ffn_hc"
+            param = match.group(3)    # "hc_fn", "hc_base", "hc_scale"
+
+            if bid is None:
+                bid = layer_idx
+
+            prefix = "hc_attn" if hc_type == "attn_hc" else "hc_ffn"
+
+            kind = {"hc_fn": "fn", "hc_base": "base", "hc_scale": "scale"}.get(param)
+            if kind is None:
+                raise ValueError(f"unexpected mHC tensor: {name}")
+            tensor_enum = self._hc_tensor_map[(prefix, kind)]
+            gguf_name = self.format_tensor_name(tensor_enum, bid)
+            yield (gguf_name, data_torch)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -58,7 +58,7 @@ def parse_args() -> argparse.Namespace:
    )
    parser.add_argument(
        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "tq1_0", "tq2_0", "auto"], default="auto",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto to preserve supported FP8 weights and otherwise select the highest-fidelity 16-bit float type",
    )
    parser.add_argument(
        "--bigendian", action="store_true",
@@ -127,7 +127,7 @@ def parse_args() -> argparse.Namespace:
    )
    parser.add_argument(
        "--dspark", action="store_true",
-        help="Export only the DeepSeek-V4 DSpark draft tensors as a separate GGUF.",
+        help="Export only the DeepSeek-V4 or Mistral Large 4 DSpark draft tensors as a separate GGUF.",
    )
    parser.add_argument(
        "--mistral-format", action="store_true",
@@ -255,6 +255,9 @@ def main() -> None:
            assert hparams.get("vision_encoder") is not None, "This model does not support multimodal"
            from conversion.pixtral import PixtralModel
            model_class = PixtralModel
+        elif hparams.get("moe") is not None and hparams.get("dsa") is not None:
+            from conversion.mistral4 import MistralLarge4Model
+            model_class = MistralLarge4Model
        elif hparams.get("moe") is not None:
            from conversion.mistral import MistralMoeModel
            model_class = MistralMoeModel
@@ -267,10 +270,16 @@ def main() -> None:
            sys.exit(1)

        if args.dspark:
-            if is_mistral_format or model_architecture not in ("DeepseekV4ForCausalLM", "DeepseekV41ForCausalLM"):
-                logger.error("--dspark is only supported for DeepseekV4ForCausalLM and DeepseekV41ForCausalLM")
+            if is_mistral_format or model_architecture == "MistralLarge4ForConditionalGeneration":
+                if model_class.__name__ != "MistralLarge4Model":
+                    logger.error("--dspark is only supported for Mistral Large 4 among Mistral-format models")
+                    sys.exit(1)
+                from conversion.mistral4 import MistralLarge4DSparkModel
+                model_class = MistralLarge4DSparkModel
+            elif model_architecture not in ("DeepseekV4ForCausalLM", "DeepseekV41ForCausalLM"):
+                logger.error("--dspark is only supported for DeepseekV4ForCausalLM, DeepseekV41ForCausalLM and Mistral Large 4")
                sys.exit(1)
-            if model_architecture == "DeepseekV41ForCausalLM":
+            elif model_architecture == "DeepseekV41ForCausalLM":
                from conversion.deepseek41 import DeepseekV41DSparkModel
                model_class = DeepseekV41DSparkModel
            else:
@@ -198,6 +198,7 @@ pre_computed_hashes = [
    # jina-v2-de variants
    {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/aari1995/German_Semantic_V3", "chkhsh": "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df"},
    {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/evilfreelancer/ruGPT3XL", "chkhsh": "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4"},
+    {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Lernex/Metis-1.6-Logos", "chkhsh": "df1bc34a6366466bcb36a9a1b1cc0afdb08cfe21c9250602c930f159085390a9"},
    # lfm2 variants
    {"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B", "chkhsh": "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7"},
    # hrm-text (DFM Mimir) is SPM-style BPE: normalizer maps ' ' -> '▁', merges
@@ -161,7 +161,7 @@ lspci -n | grep -Eo '19e5:d[0-9a-f]{3}' | cut -d: -f2
 | FP16                   | Support | Support |
 | Q8_0                   | Support | Partial |
 | Q4_0                   | Support | Partial |
-| BF16                   | Support |         |
+| BF16                   | Support |    No   |

 > **310P note**
 > - `Q8_0`: data transform / buffer path is implemented, and `GET_ROWS` is supported, but quantized `MUL_MAT` / `MUL_MAT_ID` are not supported.
@@ -274,7 +274,9 @@ cmake --build build --config release

    **Notes**:

-      - CANN backend only supports FP16/Q4_0/Q8_0 models currently.
+      - On 910B the CANN backend supports FP16, BF16 and Q4_0/Q8_0 models. On 310P, BF16 is not
+        supported and quantized `MUL_MAT` / `MUL_MAT_ID` are not implemented, so only FP16 models
+        are supported (see [DataType Supports](#datatype-supports)).

 2. **Launch inference**

@@ -355,3 +357,7 @@ Enable ACL graph execution during the prefill stage, default is false. This opti
 ### GGML_CANN_OPERATOR_FUSION

 Enable operator fusion during computation, default is false. This option fuses compatible operators (e.g., ADD + RMS_NORM) to reduce overhead and improve performance.
+
+### GGML_CANN_NO_PINNED
+
+Disable pinned (host-locked) host buffers. When it is set (any value), `aclrtMallocHost` is not called: the CANN host buffer type allocates regular CPU memory instead and the backend reports no `host_buffer` capability, so host-side staging uses pageable memory. Use it when pinned host allocation fails, or when the pinned memory is needed by other processes on the host.
@@ -325,10 +325,14 @@ on 4 physical NPUs, or `--devices 'HTP0[0-1:0],HTP1[0-1:1]'` on 2 physical NPUs
  Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).

 - `GGML_HEXAGON_VTCM_SIZE=0`
-  Overrides the VTCM size used by each session. Accepts `K`, `M` and `G` suffixes; a plain number is in MB. The default (0) uses the size the hardware reports. Values above the hardware size will fail to start the session.
+  Overrides the VTCM size used by each session. Accepts `K`, `M` and `G` suffixes; a plain number is in MB.
+  The default (0) uses the size the hardware reports. Values above the hardware size will fail to start the session.

- `GGML_HEXAGON_HOSTBUF=1` (default: 0, disabled)
-  Enables allocating host buffers for debugging. By default, host buffers are disabled.
+- `GGML_HEXAGON_HOSTBUF=0` (default: 1, enabled)
+  Controls host buffer (`HTP*-HOST`) allocation in shared FastRPC memory to avoid host-device tensor copies.
+  CPU-assigned weights (input embeddings like `token_embd` and `per_layer_token_embd`, and tensors overridden with `-ot ...=CPU`)
+  are placed in host buffers as well. On older drivers/FW, large models can run out of IOVA space when these are mapped.
+  Pass `--no-host` to keep CPU-assigned weights in regular CPU memory (`-ot ...=CPU` alone is not sufficient).

 - `GGML_HEXAGON_DMA64=0` (default: enabled on v81+)
  Disables 64-bit DMA for model weights. Set to `1` to enable it explicitly on a supported architecture.
@@ -249,7 +249,7 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 | MXFP4      | ✅           | ❓    | ❓     |
 | Q5_0       | ✅           | ❓    | ❓     |
 | Q5_1       | ✅           | ❓    | ❓     |
-| Q8_0       | ✅           | ❓    | ❓     |
+| Q8_0       | ✅           | ✅    | ❓     |
 | Q2_K       | 🚫           | ❓    | ❓     |
 | Q3_K       | ✅           | ❓    | ❓     |
 | Q4_K       | ✅           | ❓    | ❓     |
@@ -266,11 +266,11 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 | IQ1_M      | 🚫           | ❓    | ❓     |
 | IQ4_NL     | ✅           | ❓    | ❓     |
 | IQ4_XS     | ✅           | ❓    | ❓     |
-| FP32->FP16 | 🚫           | ❓    | ❓     |
+| FP32->FP16 | ✅           | ❓    | ❓     |
 | FP16->FP32 | 🚫           | ❓    | ❓     |

 -   ✅ - acceleration available
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself

-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Sep 8, 2026.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Nov 3, 2026.
@@ -481,7 +481,7 @@ cmake --build build --config Release
 Now you can load the model in conversation mode using `Vulkan`

 ```sh
-build/bin/Release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2 -cnv
+build/bin/Release/llama-cli -m "[PATH TO MODEL]" -ngl 100 -c 16384 -t 10 -n -2
 ```

 **MSYS2**
@@ -0,0 +1,26 @@
+# MoE cache
+
+When the experts of MoE layers stay in host memory (`-cmoe`, `-ncmoe`, `-ot`), the MoE cache keeps the most used ones in device memory for small batches (decode, speculative decoding). Large batches still upload the experts they use.
+
+| Argument | |
+|---|---|
+| `--moe-cache-mib N` | cache size in MiB, split among the GPUs like the layers (default: 0, disabled) |
+| `--moe-cache-cpu-share F` | share of the cache misses computed on the CPU from the host experts instead of uploaded, once the cache is full (default: 0) |
+
+The cache hits run on the device. With `--moe-cache-cpu-share F` the CPU computes a share `F` of the misses while the device computes the hits and uploads the other misses into the cache. This is faster when the CPU reads the host experts faster than the device can upload them, for example with fast RAM, a PCIe 4 link or experts that are not in pinned memory. The host experts are pinned with `-lm none`, `-lm mlock` and `-lm dio`, unless `--no-host`, but not when they are mmap'd (the default). The device uploads pinned experts at the full speed of the link, so the CPU share often does not help then. Until the cache is full every miss is uploaded, so the cache fills as fast as without the CPU share. The best share depends on the machine, the model and the cache size.
+
+## Picking the arguments
+
+Put the script in front of your server command:
+
+```bash
+python3 scripts/moe-cache-tune.py -- build/bin/llama-server -m model.gguf -cmoe -t 8 -c 32768
+```
+
+It starts the server 6 times, at two cache sizes × shares 0, 0.5 and 0.95, and times the generation of a few thousand tokens. That takes a few minutes, mostly model loading.
+
+From these runs it fits the upload and CPU speeds of the machine, then replays the routing of all the runs through a copy of the cache logic for other sizes and shares. It prints the measured speeds, the predicted t/s and the recommended arguments: the lowest share within 5% of the fastest.
+
+- **Server command:** keep your own arguments (`-t`, `-c`, expert placement, load mode); the speeds and the free device memory depend on them.
+- **`--max-mib N`:** tunes a smaller cache than what fits on the device. The runs in `--out` are reused while the command stays the same, so trying another value is instant.
+- **`--markdown`:** prints Markdown tables to paste in a PR.
@@ -14,22 +14,22 @@ Legend:

 | Operation | BLAS | CANN | CPU | CUDA | ET | HTP | MTL | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
 |-----------|------|------|------|------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             ADD1 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        COL2IM_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                       CONV_1D_DW | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ |
+|                       CONV_1D_DW | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
 |                          CONV_2D | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -38,49 +38,49 @@ Legend:
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                     DSV4_HC_COMB | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                     DSV4_HC_POST | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                      DSV4_HC_PRE | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | 🟡 | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                LIGHTNING_INDEXER | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                LIGHTNING_INDEXER | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             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 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                  MUL_MAT_ID_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                  MUL_MAT_ID_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     MUL_MAT_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     MUL_MAT_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -89,45 +89,46 @@ Legend:
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROPE | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                         SET_ROWS | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SUM | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                     SWIGLU_CLAMP | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                     SWIGLU_CLAMP | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                       TOP_K_MASK | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            XIELU | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -12278,13 +12278,23 @@
 "WebGPU: WebGPU","SUM","type=f32,ne=[11,5,6,3],permute=[0,2,1,3]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","SUM","type=f32,ne=[11,5,6,3],permute=[0,3,2,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","SUM","type=f32,ne=[11,5,6,3],permute=[0,1,3,2]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","MEAN","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","MEAN","type=f32,ne=[33,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","MEAN","type=f32,ne=[33,256,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","MEAN","type=f32,ne=[32769,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","MEAN","type=f32,ne=[32,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","MEAN","type=f32,ne=[32,256,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","MEAN","type=f32,ne=[32768,1,1,1]","support","0","no","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[10,5,4,3],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[33,1,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[33,256,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[32769,1,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[32,1,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[32,256,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[32768,1,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[11,5,6,3],permute=1,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[11,5,6,3],permute=0,slice=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[11,5,6,3],permute=1,slice=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[1,7,3,2],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[1023,5,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[1024,5,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[1025,5,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[17,65537,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[512,1,1,1],permute=0,slice=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","MEAN","type=f32,ne=[281,1536,1,1],permute=0,slice=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","SUM","type=f32,ne=[33,1,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","SUM","type=f32,ne=[33,1024,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","SUM","type=f32,ne=[33,256,1,1]","support","1","yes","WebGPU"
@@ -12299,8 +12309,9 @@
 "WebGPU: WebGPU","SUM_ROWS","type=f32,ne=[33,1,1,1],permute=0,slice=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","SUM_ROWS","type=f32,ne=[33,1024,1,1],permute=0,slice=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","SUM_ROWS","type=f32,ne=[33,256,1,1],permute=0,slice=0","support","1","yes","WebGPU"
-"WebGPU: WebGPU","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","0","no","WebGPU"
-"WebGPU: WebGPU","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","0","no","WebGPU"
+"WebGPU: WebGPU","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","1","yes","WebGPU"
+"WebGPU: WebGPU","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","1","yes","WebGPU"
+"WebGPU: WebGPU","GROUP_NORM","type=f32,ne=[3,1,1,65537],num_groups=1,eps=0.000001","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1],stride_dim=-1","support","0","no","WebGPU"
 "WebGPU: WebGPU","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[256,16,2,3],stride_dim=-1","support","0","no","WebGPU"
 "WebGPU: WebGPU","ACC","type=f32,ne_a=[256,17,2,3],ne_b=[128,16,2,3],stride_dim=-1","support","0","no","WebGPU"
@@ -12326,8 +12337,16 @@
 "WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","0","no","WebGPU"
 "WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","0","no","WebGPU"
 "WebGPU: WebGPU","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","0","no","WebGPU"
-"WebGPU: WebGPU","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","0","no","WebGPU"
-"WebGPU: WebGPU","ARANGE","type=f32,start=0.000000,stop=1048576.000000,step=1.000000","support","0","no","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=3.000000,stop=3.500000,step=1.000000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=-10.000000,stop=10.000000,step=0.300000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=-1.000000,stop=62.750000,step=0.250000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=-1.000000,stop=63.000000,step=0.250000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=-1.000000,stop=63.250000,step=0.250000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=-1.000000,stop=254.750000,step=0.250000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=-1.000000,stop=255.000000,step=0.250000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=-1.000000,stop=255.250000,step=0.250000","support","1","yes","WebGPU"
+"WebGPU: WebGPU","ARANGE","type=f32,start=0.000000,stop=1048576.000000,step=1.000000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TIMESTEP_EMBEDDING","type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000","support","0","no","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","CUMSUM","type=f32,ne=[10,5,4,3]","support","1","yes","WebGPU"
@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)

 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 28)
+set(GGML_VERSION_MINOR 29)
 set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

@@ -70,7 +70,8 @@ extern "C" {
    GGML_API void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);

    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy  (const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy_n(const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors);

    //
    // Backend (stream)
@@ -95,6 +96,8 @@ extern "C" {
    GGML_API void ggml_backend_tensor_get   (const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
    GGML_API void ggml_backend_tensor_set_2d(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
    GGML_API void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+    GGML_API void ggml_backend_tensor_set_n (      struct ggml_tensor * const * tensors, const void * data, const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors);
+    GGML_API void ggml_backend_tensor_get_n (const struct ggml_tensor * const * tensors,       void * data, const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors);
    GGML_API void ggml_backend_tensor_memset(      struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);

    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
@@ -208,6 +211,18 @@ extern "C" {

    // Context management and operations for faster communication between backends, used for tensor parallelism (meta backend)
    typedef void * (*ggml_backend_comm_init_t)(ggml_backend_t * backends, size_t n_backends);
+
+    // Builds the communicator for the case where each rank is a separate process, so a process
+    // owns one backend of the communicator rather than all of them.
+    //
+    // The caller provides bootstrap_callback, which moves bytes between the ranks before the
+    // communicator exists. Each rank passes size bytes in send_buf, and recv_buf receives
+    // world_size * size bytes in rank order. The backend decides what to exchange and how many
+    // times to call it, and every rank must make the same calls in the same order.
+    // comm_init_rank blocks until all world_size ranks have joined.
+    typedef bool   (*ggml_backend_comm_bootstrap_callback)(void * user_data, const void * send_buf, void * recv_buf, size_t size);
+    typedef void * (*ggml_backend_comm_init_rank_t)(ggml_backend_t backend, int rank, int world_size, ggml_backend_comm_bootstrap_callback bootstrap_callback, void * bootstrap_callback_data);
+
    typedef void   (*ggml_backend_comm_free_t)(void * comm_ctx);
    typedef bool   (*ggml_backend_comm_allreduce_tensor_t)(void * comm_ctx, struct ggml_tensor ** tensors);

@@ -381,6 +381,8 @@ extern "C" {
    GGML_API void        ggml_fp32_to_bf16_row_ref(const float *, ggml_bf16_t *, int64_t);
    GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);

+    typedef struct { uint8_t bits; } ggml_fp8_e4m3_t;
+
    struct ggml_object;
    struct ggml_context;
    struct ggml_cgraph;
@@ -430,7 +432,8 @@ extern "C" {
        GGML_TYPE_NVFP4   = 40, // NVFP4 (4 blocks, E4M3 scale)
        GGML_TYPE_Q1_0    = 41,
        GGML_TYPE_Q2_0    = 42,
-        GGML_TYPE_COUNT   = 43,
+        GGML_TYPE_F8_E4M3 = 43,
+        GGML_TYPE_COUNT   = 44,
    };

    // [TAG_GGML_PREC]
@@ -486,6 +489,7 @@ extern "C" {
        GGML_FTYPE_MOSTLY_NVFP4   = 26, // except 1d tensors
        GGML_FTYPE_MOSTLY_Q1_0    = 27, // except 1d tensors
        GGML_FTYPE_MOSTLY_Q2_0    = 28, // except 1d tensors
+        GGML_FTYPE_MOSTLY_F8_E4M3 = 29, // except 1d tensors
    };

    // available tensor operations:
@@ -1515,6 +1519,7 @@ extern "C" {
    // note: this op has undefined behaviour if a given `a` is referenced more than once for a given `b`
    //       (i.e. duplicate experts per token are not guaranteed to be supported in all backends)
    // TODO: add backend support: https://github.com/ggml-org/llama.cpp/pull/29998#issuecomment-6153742533
+    // the CPU backend skips negative ids, their rows of the result are zero
    GGML_API struct ggml_tensor * ggml_mul_mat_id(
            struct ggml_context * ctx,
            struct ggml_tensor  * as,
@@ -2030,10 +2035,6 @@ extern "C" {
            float                 beta_slow),
        "use ggml_rope_ext_inplace instead");

-    // compute correction dims for YaRN RoPE scaling
-    GGML_API void ggml_rope_yarn_corr_dims(
-        int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
-
    // rotary position embedding backward, i.e compute dx from dy
    // a - dy
    GGML_API struct ggml_tensor * ggml_rope_ext_back(
@@ -2076,6 +2077,11 @@ extern "C" {
            struct ggml_tensor  * a,
            int                   n_offs);

+    // control rounding of YaRN correction bounds (default: true)
+    GGML_API struct ggml_tensor * ggml_rope_set_truncate(
+            struct ggml_tensor * a,
+            bool                 truncate);
+
    // im2col
    // converts data into a format that effectively results in a convolution when combined with matrix multiplication
    GGML_API struct ggml_tensor * ggml_im2col(
@@ -2512,6 +2518,24 @@ extern "C" {
            float                 max_bias,
            float                 logit_softcap);

+    // same as ggml_flash_attn_ext, but each q slice attends to its own list of K/V rows
+    // q:       [n_embd_k, n_batch, n_head,    ne3]
+    // k:       [n_embd_k, n_rows,  n_head_kv, 1  ] shared by all q slices
+    // v:       [n_embd_v, n_rows,  n_head_kv, 1  ]
+    // mask:    [n_kv,     n_batch, ne32,      ne3]
+    // kv_rows: [n_kv,     ne3] I32, mask column i of slice i3 uses row kv_rows[i, i3] of k and v
+    // negative rows are padding, their mask entries must be -inf
+    GGML_API struct ggml_tensor * ggml_flash_attn_ext_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * q,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * mask,
+            struct ggml_tensor  * kv_rows,
+            float                 scale,
+            float                 max_bias,
+            float                 logit_softcap);
+
    GGML_DEPRECATED(GGML_API void ggml_flash_attn_ext_set_prec(
            struct ggml_tensor * a,
            enum ggml_prec       prec),
@@ -47,6 +47,9 @@ extern "C" {
    // Backend buffer
    //

+    // TODO: remove `set_tensor`, `get_tensor`, `set_tensor_2d`, `get_tensor_2d` and `cpy_tensor` from the buffer interface
+    //       they are superseded by the `set_tensor_n`, `get_tensor_n`, `cpy_tensor_n` interface
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/30881#discussion_r4400299490
    struct ggml_backend_buffer_i {
        // (optional) free the buffer
        void         (*free_buffer)  (ggml_backend_buffer_t buffer);
@@ -68,6 +71,11 @@ extern "C" {
        void         (*clear)        (ggml_backend_buffer_t buffer, uint8_t value);
        // (optional) reset any internal state due to tensor initialization, such as tensor extras
        void         (*reset)        (ggml_backend_buffer_t buffer);
+        // (optional) batched tensor data copies
+        void         (*set_tensor_n) (ggml_backend_buffer_t buffer,       struct ggml_tensor * const * tensors, const void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors);
+        void         (*get_tensor_n) (ggml_backend_buffer_t buffer, const struct ggml_tensor * const * tensors,       void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors);
+        // (optional) batched tensor copies: returns true if all copies were performed, false if unsupported and no copies were performed
+        bool         (*cpy_tensor_n) (ggml_backend_buffer_t buffer, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors);
    };

    struct ggml_backend_buffer {
@@ -489,6 +489,32 @@ static struct ggml_tensor * ggml_backend_meta_buffer_simple_tensor(const struct
    return it->second[index];
 }

+// bytes spanned by all dims except skip_dim
+static size_t ggml_backend_meta_inner_extent(const struct ggml_tensor * tensor, int skip_dim) {
+    size_t extent = 0;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (i != skip_dim && tensor->ne[i] > 1) {
+            extent += (size_t) (tensor->ne[i] - 1) * tensor->nb[i];
+        }
+    }
+    return extent;
+}
+
+// the view dim that holds the split axis of src, or -1
+// all other dims of src and view must stay inside one slab of that axis
+static int ggml_backend_meta_view_slab_dim(const struct ggml_tensor * view, const struct ggml_tensor * src, int axis) {
+    const size_t slab = src->nb[axis];
+    if (ggml_backend_meta_inner_extent(src, axis) >= slab) {
+        return -1;
+    }
+    for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
+        if (view->ne[dim] == src->ne[axis] && view->nb[dim] == slab) {
+            return ggml_backend_meta_inner_extent(view, dim) < slab ? dim : -1;
+        }
+    }
+    return -1;
+}
+
 static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(const struct ggml_tensor * tensor, bool assume_sync);

 static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
@@ -700,9 +726,39 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
            }
            GGML_ABORT("fatal error");
        }
+        // permuted view: split the view dim that spans the src split axis, other dims must be inside one element of it or outside the span
+        if (ggml_is_permuted(tensor) && !ggml_is_permuted(tensor->src[0]) && axis >= 0 && axis < GGML_MAX_DIMS-1) {
+            const size_t nb_axis = tensor->src[0]->nb[axis];
+            const size_t nb_next = tensor->src[0]->nb[axis+1];
+            for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
+                if (tensor->ne[dim] == 1 || tensor->nb[dim] < nb_axis || tensor->nb[dim]*tensor->ne[dim] != nb_next) {
+                    continue;
+                }
+                bool others_ok = true;
+                for (int i = 0; i < GGML_MAX_DIMS; i++) {
+                    if (i == dim || tensor->ne[i] == 1) {
+                        continue;
+                    }
+                    others_ok = others_ok && (tensor->nb[i]*tensor->ne[i] <= tensor->nb[dim] || tensor->nb[i] % nb_next == 0);
+                }
+                if (others_ok) {
+                    return {ggml_backend_meta_split_axis(dim), {0}, {1}, 1};
+                }
+            }
+        }
        if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED || src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
            return src_ss[0];
        }
+        // a view that only regroups dims inside one slab of the split axis keeps the split,
+        // the per-device strides are only rescaled above nb[axis] (e.g. the KV streams of a permuted q)
+        if (axis >= 0 && axis < GGML_MAX_DIMS && tensor->view_offs == tensor->src[0]->view_offs) {
+            const int dim = ggml_backend_meta_view_slab_dim(tensor, tensor->src[0], axis);
+            if (dim >= 0) {
+                ggml_backend_meta_split_state ret = src_ss[0];
+                ret.axis = ggml_backend_meta_split_axis(dim);
+                return ret;
+            }
+        }
        GGML_ABORT("view of permuted tensor not implemented");
        //return {GGML_BACKEND_SPLIT_AXIS_UNKNOWN, {0}, {1}, 1};
    };
@@ -1000,8 +1056,13 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
                split_state = handle_pad(src_ss);
            } break;
            case GGML_OP_PAD_REFLECT_1D:
-            case GGML_OP_ROLL:
-            case GGML_OP_ARANGE:
+            case GGML_OP_ROLL: {
+                split_state = handle_generic(src_ss, /*scalar_only =*/ true);
+            } break;
+            case GGML_OP_ARANGE: {
+                // no sources, every device computes the full range
+                split_state = {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
+            } break;
            case GGML_OP_TIMESTEP_EMBEDDING: {
                split_state = handle_generic(src_ss, /*scalar_only =*/ true);
            } break;
@@ -1699,6 +1760,9 @@ static const ggml_backend_buffer_i ggml_backend_meta_buffer_iface = {
    /* .cpy_tensor      = */ nullptr,
    /* .clear           = */ ggml_backend_meta_buffer_clear,
    /* .reset           = */ ggml_backend_meta_buffer_reset,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };

 bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf) {
@@ -1744,13 +1808,14 @@ static ggml_backend_buffer_t ggml_backend_meta_buffer_type_alloc_buffer_n(ggml_b
    const size_t n_simple_bufts = ggml_backend_meta_buft_n_bufts(buft);

    constexpr size_t compute_headroom = 16; // Maximum number of views per statically allocated tensor that can be created between evals.
+    constexpr size_t compute_headroom_min = 256; // A buffer with few tensors can still get many views, e.g. a KV cache shared by many layers.
    const ggml_init_params params_static = {
        /*.mem_size   =*/ n_tensors * ggml_tensor_overhead(),
        /*.mem_buffer =*/ nullptr,
        /*.no_alloc   =*/ true,
    };
    const ggml_init_params params_compute = {
-        /*.mem_size   =*/ compute_headroom * n_tensors * ggml_tensor_overhead(),
+        /*.mem_size   =*/ std::max(compute_headroom * n_tensors, compute_headroom_min) * ggml_tensor_overhead(),
        /*.mem_buffer =*/ nullptr,
        /*.no_alloc   =*/ true,
    };
@@ -402,6 +402,62 @@ size_t ggml_backend_get_max_size(ggml_backend_t backend) {
    return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
 }

+template <typename cb_get_buffer>
+static void ggml_backend_tensor_buft_groups(size_t n_tensors, cb_get_buffer && get_buffer, std::vector<std::vector<size_t>> & groups) {
+    std::unordered_map<ggml_backend_buffer_type_t, size_t> buft_idx;
+    groups.clear();
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t buf = get_buffer(i);
+        auto it = buft_idx.find(buf->buft);
+        if (it == buft_idx.end()) {
+            buft_idx[buf->buft] = groups.size();
+            groups.emplace_back();
+        }
+        groups[buft_idx[buf->buft]].push_back(i);
+    }
+}
+
+static bool ggml_backend_tensor_batch_is_2d(
+        const struct ggml_tensor * const * tensors,
+        const size_t * offsets,
+        const size_t * sizes,
+        const size_t * data_offsets,
+        size_t n_tensors,
+        size_t * base_offset,
+        size_t * size,
+        size_t * stride_tensor,
+        size_t * stride_data) {
+    if (n_tensors <= 1) {
+        return false;
+    }
+
+    const struct ggml_tensor * tensor = tensors[0];
+    const size_t size0 = sizes[0];
+
+    if (size0 == 0 || offsets[1] < offsets[0] || data_offsets[1] < data_offsets[0]) {
+        return false;
+    }
+
+    const size_t stride_tensor0 = offsets[1] - offsets[0];
+    const size_t stride_data0   = data_offsets[1] - data_offsets[0];
+
+    for (size_t i = 1; i < n_tensors; i++) {
+        if (tensors[i] != tensor || sizes[i] != size0) {
+            return false;
+        }
+        if (offsets[i] != offsets[0] + i*stride_tensor0 || data_offsets[i] != data_offsets[0] + i*stride_data0) {
+            return false;
+        }
+    }
+
+    *base_offset   = offsets[0];
+    *size          = size0;
+    *stride_tensor = stride_tensor0;
+    *stride_data   = stride_data0;
+
+    return true;
+}
+
 void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
    GGML_ASSERT(backend);
    GGML_ASSERT(tensor);
@@ -546,6 +602,154 @@ void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor, void * data,
    buf->iface.get_tensor_2d(buf, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
 }

+void ggml_backend_tensor_set_n(struct ggml_tensor * const * tensors, const void * data,
+        const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors) {
+    if (n_tensors == 0) {
+        return;
+    }
+
+    std::vector<const void *> data_ptrs(n_tensors);
+    for (size_t i = 0; i < n_tensors; i++) {
+        data_ptrs[i] = (const char *) data + data_offsets[i];
+    }
+
+    ggml_backend_buffer_t buf = NULL;
+    bool same_buft = true;
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        GGML_ASSERT(tensors[i] != NULL);
+        ggml_backend_buffer_t tensor_buf = tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        GGML_ASSERT(tensor_buf != NULL && "tensor buffer not set");
+        GGML_ASSERT(tensors[i]->data != NULL && "tensor not allocated");
+        GGML_ASSERT(offsets[i] + sizes[i] <= ggml_nbytes(tensors[i]) && "tensor write out of bounds");
+
+        if (buf == NULL) {
+            buf = tensor_buf;
+        } else if (buf->buft != tensor_buf->buft) {
+            same_buft = false;
+        }
+    }
+
+    if (same_buft && buf->iface.set_tensor_n != NULL) {
+        buf->iface.set_tensor_n(buf, tensors, data_ptrs.data(), offsets, sizes, n_tensors);
+        return;
+    }
+
+    if (!same_buft) {
+        std::vector<std::vector<size_t>> groups;
+        ggml_backend_tensor_buft_groups(n_tensors, [&](size_t i) {
+            return tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        }, groups);
+
+        for (size_t g = 0; g < groups.size(); g++) {
+            std::vector<ggml_tensor *> g_tensors;
+            std::vector<size_t> g_offsets;
+            std::vector<size_t> g_sizes;
+            std::vector<size_t> g_data_offsets;
+            g_tensors.reserve(groups[g].size());
+            g_offsets.reserve(groups[g].size());
+            g_sizes.reserve(groups[g].size());
+            g_data_offsets.reserve(groups[g].size());
+
+            for (size_t i : groups[g]) {
+                g_tensors.push_back(tensors[i]);
+                g_offsets.push_back(offsets[i]);
+                g_sizes.push_back(sizes[i]);
+                g_data_offsets.push_back(data_offsets[i]);
+            }
+
+            ggml_backend_tensor_set_n(g_tensors.data(), data, g_offsets.data(), g_sizes.data(), g_data_offsets.data(), g_tensors.size());
+        }
+        return;
+    }
+
+    size_t base_offset = 0, size = 0, stride_tensor = 0, stride_data = 0;
+    if (ggml_backend_tensor_batch_is_2d(tensors, offsets, sizes, data_offsets, n_tensors,
+                &base_offset, &size, &stride_tensor, &stride_data)) {
+        ggml_backend_tensor_set_2d(tensors[0], data_ptrs[0], base_offset, size,
+                n_tensors, stride_tensor, stride_data);
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_tensor_set(tensors[i], data_ptrs[i], offsets[i], sizes[i]);
+    }
+}
+
+void ggml_backend_tensor_get_n(const struct ggml_tensor * const * tensors, void * data,
+        const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors) {
+    if (n_tensors == 0) {
+        return;
+    }
+
+    std::vector<void *> data_ptrs(n_tensors);
+    for (size_t i = 0; i < n_tensors; i++) {
+        data_ptrs[i] = (char *) data + data_offsets[i];
+    }
+
+    ggml_backend_buffer_t buf = NULL;
+    bool same_buft = true;
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        GGML_ASSERT(tensors[i] != NULL);
+        ggml_backend_buffer_t tensor_buf = tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        GGML_ASSERT(tensor_buf != NULL && "tensor buffer not set");
+        GGML_ASSERT(tensors[i]->data != NULL && "tensor not allocated");
+        GGML_ASSERT(offsets[i] + sizes[i] <= ggml_nbytes(tensors[i]) && "tensor read out of bounds");
+
+        if (buf == NULL) {
+            buf = tensor_buf;
+        } else if (buf->buft != tensor_buf->buft) {
+            same_buft = false;
+        }
+    }
+
+    if (same_buft && buf->iface.get_tensor_n != NULL) {
+        buf->iface.get_tensor_n(buf, tensors, data_ptrs.data(), offsets, sizes, n_tensors);
+        return;
+    }
+
+    if (!same_buft) {
+        std::vector<std::vector<size_t>> groups;
+        ggml_backend_tensor_buft_groups(n_tensors, [&](size_t i) {
+            return tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        }, groups);
+
+        for (size_t g = 0; g < groups.size(); g++) {
+            std::vector<const ggml_tensor *> g_tensors;
+            std::vector<size_t> g_offsets;
+            std::vector<size_t> g_sizes;
+            std::vector<size_t> g_data_offsets;
+            g_tensors.reserve(groups[g].size());
+            g_offsets.reserve(groups[g].size());
+            g_sizes.reserve(groups[g].size());
+            g_data_offsets.reserve(groups[g].size());
+
+            for (size_t i : groups[g]) {
+                g_tensors.push_back(tensors[i]);
+                g_offsets.push_back(offsets[i]);
+                g_sizes.push_back(sizes[i]);
+                g_data_offsets.push_back(data_offsets[i]);
+            }
+
+            ggml_backend_tensor_get_n(g_tensors.data(), data, g_offsets.data(), g_sizes.data(), g_data_offsets.data(), g_tensors.size());
+        }
+        return;
+    }
+
+    size_t base_offset = 0, size = 0, stride_tensor = 0, stride_data = 0;
+    if (ggml_backend_tensor_batch_is_2d(tensors, offsets, sizes, data_offsets, n_tensors,
+                &base_offset, &size, &stride_tensor, &stride_data)) {
+        ggml_backend_tensor_get_2d(tensors[0], data_ptrs[0], base_offset, size,
+                n_tensors, stride_tensor, stride_data);
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_tensor_get(tensors[i], data_ptrs[i], offsets[i], sizes[i]);
+    }
+}
+
 void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
    GGML_ASSERT(tensor);
    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -648,6 +852,59 @@ void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor
    }
 }

+void ggml_backend_tensor_copy_n(const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors) {
+    if (n_tensors == 0) {
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; ++i) {
+        GGML_ASSERT(ggml_are_same_layout(srcs[i], dsts[i]) && "cannot copy tensors with different layouts");
+    }
+
+    ggml_backend_buffer_t buf = dsts[0]->view_src ? dsts[0]->view_src->buffer : dsts[0]->buffer;
+    bool same_buft = true;
+    for (size_t i = 0; i < n_tensors; ++i) {
+        ggml_backend_buffer_t dst_buf = dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        if (dst_buf->buft != buf->buft) {
+            same_buft = false;
+            break;
+        }
+    }
+
+    // if cpy_tensor_n returns false, it must not have performed any copies
+    if (same_buft && buf->iface.cpy_tensor_n != NULL) {
+        if (buf->iface.cpy_tensor_n(buf, srcs, dsts, n_tensors)) {
+            return;
+        }
+    }
+
+    if (!same_buft) {
+        std::vector<std::vector<size_t>> groups;
+        ggml_backend_tensor_buft_groups(n_tensors, [&](size_t i) {
+            return dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        }, groups);
+
+        for (size_t g = 0; g < groups.size(); g++) {
+            std::vector<const ggml_tensor *> g_srcs;
+            std::vector<ggml_tensor *> g_dsts;
+            g_srcs.reserve(groups[g].size());
+            g_dsts.reserve(groups[g].size());
+
+            for (size_t i : groups[g]) {
+                g_srcs.push_back(srcs[i]);
+                g_dsts.push_back(dsts[i]);
+            }
+
+            ggml_backend_tensor_copy_n(g_srcs.data(), g_dsts.data(), g_srcs.size());
+        }
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; ++i) {
+        ggml_backend_tensor_copy(srcs[i], dsts[i]);
+    }
+}
+
 void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst) {
    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");

@@ -853,6 +1110,9 @@ static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
    /* .cpy_tensor      = */ NULL,
    /* .clear           = */ ggml_backend_multi_buffer_clear,
    /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };

 ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
@@ -986,6 +1246,14 @@ struct ggml_backend_sched {
    int prev_graph_size;

    int debug_realloc;
+
+    // inputs of the next split read from the backend of the current split before it computes, see ggml_backend_sched_prefetch_inputs
+    ggml_backend_event_t prefetch_events[GGML_SCHED_MAX_BACKENDS]; // created on first use, NULL if the backend has no events
+    bool                 prefetch_tried [GGML_SCHED_MAX_BACKENDS];
+    struct ggml_tensor * prefetched[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_prefetched;
+    int prefetch_split;     // the split that reads the prefetched inputs
+    int prefetch_backend_id; // the backend that copies them
 };

 #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
@@ -1826,6 +2094,16 @@ static void ggml_backend_sched_copy_input(ggml_backend_sched_t sched, struct ggm
    ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, input);
    struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);

+    // read before the previous split, wait for the copy only
+    if (split == &sched->splits[sched->prefetch_split]) {
+        for (int i = 0; i < sched->n_prefetched; i++) {
+            if (sched->prefetched[i] == input) {
+                ggml_backend_event_synchronize(sched->prefetch_events[sched->prefetch_backend_id]);
+                return;
+            }
+        }
+    }
+
    if (input->flags & GGML_TENSOR_FLAG_INPUT) {
        // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
        if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
@@ -1862,6 +2140,62 @@ static void ggml_backend_sched_copy_input(ggml_backend_sched_t sched, struct ggm
    }
 }

+// the inputs of the next split that are computed on the backend of this split before it, and copied to host memory,
+// are copied before this split computes, so that the next split waits for these copies instead of this split
+// e.g. a CPU split runs while a GPU split computes, if it reads only the results of the earlier splits
+static void ggml_backend_sched_prefetch_inputs(ggml_backend_sched_t sched, int split_id) {
+    sched->n_prefetched = 0;
+    if (split_id + 1 >= sched->n_splits) {
+        return;
+    }
+    struct ggml_backend_sched_split * split = &sched->splits[split_id];
+    struct ggml_backend_sched_split * next  = &sched->splits[split_id + 1];
+    const int backend_id = split->backend_id;
+    ggml_backend_t backend = sched->backends[backend_id];
+    if (backend->iface.get_tensor_async == NULL) {
+        return;
+    }
+
+    for (int i = 0; i < next->n_inputs && sched->n_prefetched < GGML_SCHED_MAX_SPLIT_INPUTS; i++) {
+        struct ggml_tensor * input = next->inputs[i];
+        if ((input->flags & GGML_TENSOR_FLAG_INPUT) || tensor_backend_id(input) != backend_id) {
+            continue;
+        }
+        struct ggml_tensor * input_cpy = tensor_copy(input, next->backend_id, sched->cur_copy);
+        if (input_cpy->buffer == NULL || !ggml_backend_buffer_is_host(input_cpy->buffer)) {
+            continue;
+        }
+
+        // not if this split computes or writes the input
+        const struct ggml_tensor * base = input->view_src ? input->view_src : input;
+        bool written = false;
+        for (int j = 0; j < split->graph.n_nodes && !written; j++) {
+            const struct ggml_tensor * node = split->graph.nodes[j];
+            written = !ggml_is_view_op(node->op) && (node == base || node->view_src == base);
+        }
+        if (written) {
+            continue;
+        }
+
+        if (!sched->prefetch_tried[backend_id]) {
+            sched->prefetch_tried[backend_id]  = true;
+            sched->prefetch_events[backend_id] = ggml_backend_event_new(backend->device);
+        }
+        if (sched->prefetch_events[backend_id] == NULL) {
+            return;
+        }
+
+        ggml_backend_tensor_get_async(backend, input, input_cpy->data, 0, ggml_nbytes(input));
+        sched->prefetched[sched->n_prefetched++] = input;
+    }
+
+    if (sched->n_prefetched > 0) {
+        sched->prefetch_split      = split_id + 1;
+        sched->prefetch_backend_id = backend_id;
+        ggml_backend_event_record(sched->prefetch_events[backend_id], backend);
+    }
+}
+
 static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
    GGML_ASSERT(sched);
    struct ggml_backend_sched_split * splits = sched->splits;
@@ -1890,6 +2224,9 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                ggml_backend_sched_copy_input(sched, split, split->inputs[input_id]);
            }
        }
+        // after the copies above, which may read host memory reused by the copies of the next split
+        // before the weights, whose copy callback may queue large uploads
+        ggml_backend_sched_prefetch_inputs(sched, split_id);
        for (int input_id = 0; input_id < split->n_inputs; input_id++) {
            if (ggml_backend_sched_is_host_weight(split->inputs[input_id])) {
                ggml_backend_sched_copy_input(sched, split, split->inputs[input_id]);
@@ -2026,6 +2363,7 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
        for (int c = 0; c < sched->n_copies; c++) {
            ggml_backend_event_free(sched->events[b][c]);
        }
+        ggml_backend_event_free(sched->prefetch_events[b]);
    }
    ggml_gallocr_free(sched->galloc);
    ggml_free(sched->ctx);
@@ -2532,6 +2870,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
    /* .clear           = */ ggml_backend_cpu_buffer_clear,
    /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };

 static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
@@ -2546,6 +2887,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
    /* .clear           = */ ggml_backend_cpu_buffer_clear,
    /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };

 // CPU backend buffer type
@@ -2934,7 +2934,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
    const float theta_scale = powf(freq_base, -2.0f / n_dims);

    float corr_dims[2];
-    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, true, corr_dims);

    bool       is_neox    = mode & GGML_ROPE_TYPE_NEOX;
    const bool is_imrope  = mode == GGML_ROPE_TYPE_IMROPE;  // qwen3vl apply interleaved mrope
@@ -3289,7 +3289,7 @@ void ggml_cann_rope_cache_preload(ggml_backend_cann_context & ctx, ggml_tensor *
    const float theta_scale = powf(freq_base, -2.0f / n_dims);

    float corr_dims[2];
-    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, true, corr_dims);

    bool       is_neox    = mode & GGML_ROPE_TYPE_NEOX;
    const bool is_imrope  = mode == GGML_ROPE_TYPE_IMROPE;
@@ -1478,6 +1478,9 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
    /* .cpy_tensor      = */ ggml_backend_cann_buffer_cpy_tensor,
    /* .clear           = */ ggml_backend_cann_buffer_clear,
    /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };

 /**
@@ -2544,6 +2547,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
            }
        case GGML_OP_ROPE:
            {
+                if (!ggml_rope_get_truncate(op)) {
+                    return false;
+                }
                if (((const int32_t *) op->op_params)[15] != 0) {
                    return false; // FIXME: support ggml_rope_set_offset
                }
@@ -2667,6 +2673,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
            return true;
        case GGML_OP_FLASH_ATTN_EXT:
            {
+                // K/V row lists (ggml_flash_attn_ext_rows) are not supported
+                if (op->src[5] != nullptr) {
+                    return false;
+                }
 #ifdef ASCEND_310P
                // FA not support on 310p device
                return false;
@@ -116,6 +116,9 @@ static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
    /* .cpy_tensor      = */ nullptr,
    /* .clear           = */ ggml_backend_amx_buffer_clear,
    /* .reset           = */ nullptr,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };

 static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
@@ -29,7 +29,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
    assert(k % QK8_0 == 0);
    const int nb = k / QK8_0;

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
    const size_t vl_calc = __riscv_vsetvl_e32m8(QK8_0);
    const size_t vl_save = __riscv_vsetvl_e64m2(4);
@@ -43,6 +43,14 @@ static inline float f32_to_f32(float x) {
    return x;
 }

+static inline float f8_e4m3_to_f32(ggml_fp8_e4m3_t x) {
+    return ggml_f8_e4m3_to_fp32(x.bits);
+}
+
+static inline ggml_fp8_e4m3_t f32_to_f8_e4m3(float x) {
+    return { ggml_fp32_to_f8_e4m3(x) };
+}
+
 // TODO - merge this into the traits table, after using row-based conversions
 template <class T>
 struct type_conversion_table;
@@ -65,6 +73,12 @@ struct type_conversion_table<ggml_bf16_t> {
    static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
 };

+template <>
+struct type_conversion_table<ggml_fp8_e4m3_t> {
+    static constexpr float (*to_f32)(ggml_fp8_e4m3_t) = f8_e4m3_to_f32;
+    static constexpr ggml_fp8_e4m3_t (*from_f32)(float) = f32_to_f8_e4m3;
+};
+
 template <>
 struct type_conversion_table<int32_t> {
    static constexpr float (*to_f32)(int32_t) = i32_to_f32;
@@ -30,7 +30,7 @@ struct ggml_compute_params {
 };


-#if defined(_MSC_VER)
+#if defined(_MSC_VER) && !defined(__clang__)

 #define m512bh(p) p
 #define m512i(p) p
@@ -345,7 +345,7 @@ static inline int32x4_t ggml_nvfp4_dot8(const int8x8_t q4_lo, const int8x8_t q8_
 #include <immintrin.h>
 #endif

-#ifdef __riscv_v_intrinsic
+#if defined(__riscv_v)
 #include <riscv_vector.h>
 #endif

@@ -86,6 +86,9 @@ float ggml_table_f32_e8m0_half[1 << 8];
 // precomputed f32 table for ue4m3 (1 KB) (simd-mappings.h)
 float ggml_table_f32_ue4m3[1 << 8];

+// Signed FP8 needs its own table: UE4M3 ignores the sign, halves values for NVFP4, and maps NaN to zero.
+float ggml_table_f32_f8_e4m3[1 << 8];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
    int sve_cnt;
@@ -296,6 +299,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
        .vec_dot_type             = GGML_TYPE_Q8_0,
        .nrows                    = 1,
    },
+    [GGML_TYPE_F8_E4M3] = {
+        .from_float               = quantize_row_f8_e4m3,
+        .vec_dot                  = ggml_vec_dot_f8_e4m3_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
    [GGML_TYPE_Q2_K] = {
        .from_float               = quantize_row_q2_K,
        .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
@@ -737,7 +746,7 @@ static void ggml_init_arm_arch_features(void) {}
 #endif
 #endif // __ARM_ARCH

-#if defined(__riscv) && defined(__riscv_v_intrinsic)
+#if defined(__riscv) && defined(__riscv_v)
 #include <riscv_vector.h>
 static void ggml_init_riscv_arch_features(void) {
    ggml_riscv_arch_features.rvv_vlen = __riscv_vlenb();
@@ -2876,7 +2885,7 @@ struct ggml_cplan ggml_graph_plan(
                case GGML_OP_ADD_ID:
                case GGML_OP_ADD1:
                    {
-                        if (ggml_is_quantized(node->src[0]->type)) {
+                        if (ggml_is_quantized(node->src[0]->type) || node->src[0]->type == GGML_TYPE_F8_E4M3) {
                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                        }
                    } break;
@@ -2924,7 +2933,7 @@ struct ggml_cplan ggml_graph_plan(
                    } break;
                case GGML_OP_OUT_PROD:
                    {
-                        if (ggml_is_quantized(node->src[0]->type) ||
+                        if (ggml_is_quantized(node->src[0]->type) || node->src[0]->type == GGML_TYPE_F8_E4M3 ||
                            node->src[0]->type == GGML_TYPE_F16) {
                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                        }
@@ -3547,7 +3556,7 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
        _mm_storeu_ps(y + i, y_vec);
    }

-#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfhmin)
+#elif defined(__riscv_v) && defined(__riscv_zvfhmin)
    // calculate step size
    const int epr = __riscv_vsetvlmax_e16m2();
    const int step = epr * 2;
@@ -3617,7 +3626,7 @@ void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) {
                                        (const __m128i *)(x + i))),
                                16)));
    }
-#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfmin)
+#elif defined(__riscv_v) && defined(__riscv_zvfbfmin)
    // calculate step size
    const int epr = __riscv_vsetvlmax_e16m2();
    const int step = epr * 2;
@@ -3737,7 +3746,7 @@ int ggml_cpu_has_arm_fma(void) {
 }

 int ggml_cpu_has_riscv_v(void) {
-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
    return 1;
 #else
    return 0;
@@ -3745,7 +3754,7 @@ int ggml_cpu_has_riscv_v(void) {
 }

 int ggml_cpu_get_rvv_vlen(void) {
-#if defined(__riscv) && defined(__riscv_v_intrinsic)
+#if defined(__riscv) && defined(__riscv_v)
    return ggml_riscv_arch_features.rvv_vlen;
 #else
    return 0;
@@ -3908,6 +3917,7 @@ void ggml_cpu_init(void) {
            // initialize UE4M3 table (256 entries)
            for (int i = 0; i < (1 << 8); ++i) {
                ggml_table_f32_ue4m3[i] = ggml_ue4m3_to_fp32(i);
+                ggml_table_f32_f8_e4m3[i] = ggml_f8_e4m3_to_fp32(i);
            }

            const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
@@ -69,7 +69,7 @@
 #define VECTOR_REGISTERS 16
 #endif

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
 #define LMUL 4
 #endif

@@ -180,7 +180,7 @@ inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
 }
 #endif

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
 template <> inline vfloat32m1_t madd(vfloat32m1_t a, vfloat32m1_t b, vfloat32m1_t c) {
    return __riscv_vfmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
 }
@@ -277,7 +277,7 @@ inline float hsum(__m512 x) {
 }
 #endif // __AVX512F__

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
 inline float hsum(vfloat32m1_t x) {
    return __riscv_vfmv_f_s_f32m1_f32(
        __riscv_vfredusum_vs_f32m1_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m1()));
@@ -458,7 +458,7 @@ template <> inline __m512bh load_partial(const ggml_bf16_t *p, int n) {
 #endif
 #endif

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
 template <> inline vfloat32m1_t load(const float *p) {
    return __riscv_vle32_v_f32m1(p, __riscv_vsetvlmax_e32m1());
 }
@@ -503,7 +503,7 @@ template <> inline vbfloat16m4_t load(const ggml_bf16_t *p) {
 }
 #endif

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
 template <typename T> T set_zero();

 template <> inline vfloat32m1_t set_zero() {
@@ -520,7 +520,7 @@ template <> inline vfloat32m8_t set_zero() {
 }
 #endif

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
 template <typename T> size_t vlmax() {
    if constexpr (std::is_same_v<T, vfloat32m1_t>) { return  __riscv_vsetvlmax_e32m1(); }
    else if constexpr (std::is_same_v<T, vfloat32m2_t>) { return  __riscv_vsetvlmax_e32m2(); }
@@ -732,7 +732,7 @@ class tinyBLAS {
    const int64_t ldc;
 };

-#if defined(__riscv_v_intrinsic)
+#if defined(__riscv_v)
 template <typename D, typename V, typename TA, typename TB, typename TC>
 class tinyBLAS_RVV {
  public:
@@ -3958,7 +3958,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)
    #if LMUL == 1
        tinyBLAS_RVV<vfloat32m1_t, vfloat32m1_t, float, float, float> tb{ params,
            k, (const float *)A, lda,
@@ -537,26 +537,36 @@ void ggml_compute_forward_dup(
    switch (src0->type) {
        case GGML_TYPE_F16:
            {
-                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_fp16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_bf16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<ggml_fp16_t, float      >(params, dst);
+                /**/ if (dst->type == GGML_TYPE_F16)     ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16)    ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)     ggml_compute_forward_dup_flt<ggml_fp16_t, float      >(params, dst);
+                else if (dst->type == GGML_TYPE_F8_E4M3) ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_fp8_e4m3_t>(params, dst);
                else ggml_compute_forward_dup_to_q<ggml_fp16_t>(params, dst);
            } break;
        case GGML_TYPE_BF16:
            {
-                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_fp16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_bf16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<ggml_bf16_t, float      >(params, dst);
+                /**/ if (dst->type == GGML_TYPE_F16)     ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16)    ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)     ggml_compute_forward_dup_flt<ggml_bf16_t, float      >(params, dst);
+                else if (dst->type == GGML_TYPE_F8_E4M3) ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_fp8_e4m3_t>(params, dst);
                else ggml_compute_forward_dup_to_q<ggml_bf16_t>(params, dst);
            } break;
        case GGML_TYPE_F32:
            {
-                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<float, ggml_fp16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<float, ggml_bf16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<float, float      >(params, dst);
-                else if (dst->type == GGML_TYPE_I32)  ggml_compute_forward_dup_flt<float, int32_t    >(params, dst);
+                /**/ if (dst->type == GGML_TYPE_F16)     ggml_compute_forward_dup_flt<float, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16)    ggml_compute_forward_dup_flt<float, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)     ggml_compute_forward_dup_flt<float, float      >(params, dst);
+                else if (dst->type == GGML_TYPE_F8_E4M3) ggml_compute_forward_dup_flt<float, ggml_fp8_e4m3_t>(params, dst);
+                else if (dst->type == GGML_TYPE_I32)     ggml_compute_forward_dup_flt<float, int32_t    >(params, dst);
                else ggml_compute_forward_dup_to_q<float>(params, dst);
            } break;
+        case GGML_TYPE_F8_E4M3:
+            {
+                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<ggml_fp8_e4m3_t, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<ggml_fp8_e4m3_t, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<ggml_fp8_e4m3_t, float      >(params, dst);
+                else GGML_ABORT("not implemented");
+            } break;
        case GGML_TYPE_I32:
            {
                if (dst->type == GGML_TYPE_F32) ggml_compute_forward_dup_flt<int32_t, float>(params, dst);
@@ -605,8 +615,9 @@ static void ggml_compute_forward_add_q_f32(
    GGML_ASSERT(nb1 <= nb2);
    GGML_ASSERT(nb2 <= nb3);

-    GGML_ASSERT(ggml_is_quantized(src0->type));
+    GGML_ASSERT(ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F8_E4M3);
    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(ne00 % ggml_blck_size(type) == 0);

    // rows per thread
    const int dr = (nr + nth - 1)/nth;
@@ -636,8 +647,6 @@ static void ggml_compute_forward_add_q_f32(
        float * src1_row = (float *)((char *) src1->data + (i11*nb11 + i12*nb12 + i13*nb13));
        void  * dst_row  = (void *) ((char *)  dst->data + ( i1*nb1  +  i2*nb2  +  i3*nb3));

-        assert(ne00 % 32 == 0);
-
        // unquantize row from src0 to temp buffer
        dequantize_row_q(src0_row, wdata, ne00);
        // add src1
@@ -673,6 +682,7 @@ void ggml_compute_forward_add(
        case GGML_TYPE_Q8_0:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -955,7 +965,7 @@ static void ggml_compute_forward_add1_q_f32(
    GGML_ASSERT(nb1 <= nb2);
    GGML_ASSERT(nb2 <= nb3);

-    GGML_ASSERT(ggml_is_quantized(src0->type));
+    GGML_ASSERT(ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F8_E4M3);
    GGML_ASSERT(dst->type == src0->type);
    GGML_ASSERT(src1->type == GGML_TYPE_F32);

@@ -1130,6 +1140,7 @@ void ggml_compute_forward_add1(
        case GGML_TYPE_Q8_1:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -1261,6 +1272,7 @@ void ggml_compute_forward_acc(
        case GGML_TYPE_Q8_1:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -4804,6 +4816,7 @@ void ggml_compute_forward_out_prod(
        case GGML_TYPE_Q8_0:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -5129,6 +5142,7 @@ void ggml_compute_forward_set(
        case GGML_TYPE_Q8_1:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -5354,6 +5368,7 @@ void ggml_compute_forward_get_rows(
        case GGML_TYPE_Q8_1:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -6114,6 +6129,7 @@ void ggml_compute_forward_clamp(
        case GGML_TYPE_Q8_1:
        case GGML_TYPE_MXFP4:
        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -6294,8 +6310,8 @@ static void ggml_compute_forward_rope_flt(
    //const int n_past     = ((int32_t *) dst->op_params)[0];
    const int n_dims     = ((int32_t *) dst->op_params)[1];
    const int mode       = ((int32_t *) dst->op_params)[2];
-    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const bool truncate = ggml_rope_get_truncate(dst);

    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
@@ -6340,7 +6356,7 @@ static void ggml_compute_forward_rope_flt(
    const float theta_scale = powf(freq_base, -2.0f/n_dims);

    float corr_dims[2];
-    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, truncate, corr_dims);

    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
@@ -8828,6 +8844,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
    const ggml_tensor * v     = dst->src[2];
    const ggml_tensor * mask  = dst->src[3];
    const ggml_tensor * sinks = dst->src[4];
+    const ggml_tensor * kv_rows = dst->src[5];

    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
@@ -8935,6 +8952,8 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
        const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3));
        q_to_vec_dot(pq, Q_q, DK);

+        const int32_t * rows = kv_rows ? (const int32_t *) ((const char *) kv_rows->data + iq3*kv_rows->nb[1]) : nullptr;
+
        // online softmax / attention
        // loop over n_kv and n_head_kv
        // ref: https://arxiv.org/pdf/2112.05682.pdf
@@ -8945,9 +8964,14 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
                continue;
            }

+            const int64_t ik1 = rows ? rows[ic] : ic;
+            if (ik1 < 0) {
+                continue;
+            }
+
            float s; // KQ value

-            const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3);
+            const char * k_data = (const char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3);
            kq_vec_dot(DK, &s, 0, k_data, 0, Q_q, 0, 1);

            s = s*scale; // scale KQ value
@@ -8963,7 +8987,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
            float ms = 1.0f; // upon new higher max val, scale VKQ and KQ sum with this value
            float vs = 1.0f; // post-softmax KQ value, expf(s - M)

-            const char * v_data = ((const char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3));
+            const char * v_data = ((const char *) v->data + (ik1*nbv1 + iv2*nbv2 + iv3*nbv3));

            if (v->type == GGML_TYPE_F16) {
                if (s > M) {
@@ -9130,6 +9154,10 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
    static constexpr int Q_TILE_SZ  = ggml_fa_tile_config::Q;
    static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV;

+    // with kv_rows the loop runs over the mask columns and reads K/V rows through kv_rows
+    const ggml_tensor * kv_rows = dst->src[5];
+    const int64_t n_kv_cols = kv_rows ? kv_rows->ne[0] : nek1;
+
    int ir = ir0;
    while (ir < ir1) {
        // q indices for the start of this tile
@@ -9137,6 +9165,10 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);

+        // padding rows are masked, read row 0 for them
+        const int32_t * rows = kv_rows ? (const int32_t *) ((const char *) kv_rows->data + iq3*kv_rows->nb[1]) : nullptr;
+        auto kv_row = [rows](int64_t i) -> int64_t { return rows ? std::max<int32_t>(rows[i], 0) : i; };
+
        // Number of valid rows in this tile:
        // - limited by tile size (Q_TILE_SZ)
        // - limited by chunk boundary (ir1 - ir)
@@ -9196,8 +9228,8 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
        memset(K_f32, 0, DK * KV_TILE_SZ * sizeof(float));
        memset(V32,   0, KV_TILE_SZ * DV * sizeof(float));

-        for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) {
-            const int kv_tile = (int)std::min((int64_t)KV_TILE_SZ, nek1 - ic);
+        for (int64_t ic = 0; ic < n_kv_cols; ic += KV_TILE_SZ) {
+            const int kv_tile = (int)std::min((int64_t)KV_TILE_SZ, n_kv_cols - ic);

            // skip the tile entirely if all the masks are -inf
            if (mask) {
@@ -9224,7 +9256,7 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
            // Pack K tile transposed: K_f32[dk][kv] so KV_TILE is contiguous (SIMD dim)
            // Zero-pad the last tile so the GEMM always operates on KV_TILE_SZ columns
            for (int tk = 0; tk < kv_tile; tk++) {
-                const char * k_data = (const char *)k->data + (ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3;
+                const char * k_data = (const char *)k->data + kv_row(ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3;
                if (kv_type == GGML_TYPE_F16) {
                    const ggml_fp16_t * k_f16 = (const ggml_fp16_t *)k_data;
                    for (int64_t dk = 0; dk < DK; dk++) {
@@ -9289,7 +9321,7 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
            // V accumulation: VKQ32 += softmax(KQ) * V
            // Pack V tile to contiguous F32, zero-padded
            for (int tk = 0; tk < kv_tile; tk++) {
-                const char * v_data = (const char *)v->data + (ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
+                const char * v_data = (const char *)v->data + kv_row(ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3;
                if (kv_type == GGML_TYPE_F16) {
                    ggml_cpu_fp16_to_fp32((const ggml_fp16_t *)v_data, V32 + tk * DV, DV);
                } else {
@@ -9461,8 +9493,12 @@ static void ggml_compute_forward_flash_attn_ext_f16(
    // When use_ref is set, force the vec-only reference implementation (no tiling, no KV-chunking)
    const bool use_ref = params->use_ref;

+    // with kv_rows the loops run over the mask columns, the split KV path does not support it
+    const ggml_tensor * kv_rows = dst->src[5];
+    const int64_t n_kv_cols = kv_rows ? kv_rows->ne[0] : nek1;
+
    const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16);
-    const bool use_split_kv_path = !use_ref && (neq1 == 1 && neq3 == 1) && kv_is_f32_or_f16 && (k->type == v->type) && q->type == GGML_TYPE_F32 && nek1 >= 512;
+    const bool use_split_kv_path = !use_ref && !kv_rows && (neq1 == 1 && neq3 == 1) && kv_is_f32_or_f16 && (k->type == v->type) && q->type == GGML_TYPE_F32 && nek1 >= 512;

    if (use_split_kv_path) {
        const int64_t chunk_size = (nek1 + nth - 1) / nth;
@@ -9541,7 +9577,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
            if (use_tiled) {
                ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
            } else {
-                ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1, 0, nek1, nullptr, 0);
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1, 0, n_kv_cols, nullptr, 0);
            }

            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
@@ -10078,7 +10114,7 @@ static void ggml_compute_forward_ssm_scan_f32(
                        }

                        sumf = GGML_F32xt_REDUCE_ONE(sum);
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
                        // todo: RVV implementation
                        const int np = 0;
    #else
@@ -11742,7 +11778,7 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
    int64_t h_stride_2d = head_size * head_size;

    #if defined(GGML_SIMD)
-        #if defined(__ARM_FEATURE_SVE) || defined(__riscv_v_intrinsic)
+        #if defined(__ARM_FEATURE_SVE) || defined(__riscv_v)
            // scalar Route to scalar implementation       //TODO: Write SVE code and RVV code
            for (int64_t t = 0; t < T; t++) {
                int64_t t_offset = t * t_stride;
@@ -3,6 +3,7 @@

 #include "ggml-cpu-impl.h"
 #include "simd-mappings.h"
+#include "vec.h"
 #include "ggml-quants.h"
 #include "quants.h"

@@ -62,6 +63,10 @@ void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, i
    quantize_row_nvfp4_ref(x, y, k);
 }

+void quantize_row_f8_e4m3(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_f8_e4m3_ref(x, y, k);
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -362,6 +367,49 @@ void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
    *s = sumf;
 }

+void ggml_vec_dot_f8_e4m3_f32(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const ggml_fp8_e4m3_t * GGML_RESTRICT x = vx;
+    const float * GGML_RESTRICT y = vy;
+
+    float sumf = 0.0f;
+    int pos = 0;
+#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
+    GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+    const int np = n - n % GGML_F32_STEP;
+    for (; pos < np; pos += GGML_F32_STEP) {
+        for (int j = 0; j < GGML_F32_ARR; ++j) {
+            const int offset = pos + j*GGML_F32_EPR;
+            GGML_F32_VEC ax;
+#if defined(__AVX512F__) && defined(__AVX512BW__)
+            const __m128i codes = _mm_loadu_si128((const __m128i *) (x + offset));
+            ax = _mm512_i32gather_ps(_mm512_cvtepu8_epi32(codes), ggml_table_f32_f8_e4m3, sizeof(float));
+#elif defined(__AVX2__) && !defined(__AVX512F__)
+            const __m128i codes = _mm_loadl_epi64((const __m128i *) (x + offset));
+            ax = _mm256_i32gather_ps(ggml_table_f32_f8_e4m3, _mm256_cvtepu8_epi32(codes), sizeof(float));
+#else
+            float values[GGML_F32_EPR];
+            for (int k = 0; k < GGML_F32_EPR; ++k) {
+                values[k] = ggml_table_f32_f8_e4m3[x[offset + k].bits];
+            }
+            ax = GGML_F32_VEC_LOAD(values);
+#endif
+            sum[j] = GGML_F32_VEC_FMA(sum[j], ax, GGML_F32_VEC_LOAD(y + offset));
+        }
+    }
+    GGML_F32_VEC_REDUCE(sumf, sum);
+#endif
+    for (; pos < n; ++pos) {
+        sumf += ggml_table_f32_f8_e4m3[x[pos].bits] * y[pos];
+    }
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    const int qk = QK8_0;
    const int nb = n / qk;
@@ -23,6 +23,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in

 void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_f8_e4m3(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);

 void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -40,6 +41,7 @@ void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y,
 // Dot product
 void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_f8_e4m3_f32(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -5,7 +5,7 @@
 #include "simd-mappings.h"

 // TODO: add support for sizeless vector types
-#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
+#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v)

 // TODO: untested on avx512
 // These are in units of GGML_F32_EPR
@@ -149,7 +149,7 @@ static void simd_gemm(
        C += N;
    }
 }
-#elif defined(GGML_SIMD) && defined(__riscv_v_intrinsic)
+#elif defined(GGML_SIMD) && defined(__riscv_v)
 // RM accumulators + 1 B vector = RM + 1 <= 8  =>  RM <= 7
 // Microkernel: C[RM x vl] += A[RM x K] * B[K x N]
 template <int RM>
@@ -14,7 +14,9 @@
 #include <arm_neon.h>
 #endif

-#if defined(__riscv_v_intrinsic)
+// __riscv_v_intrinsic reports compiler API support and can be defined without
+// enabling V; use the ISA feature macro to select RVV code.
+#if defined(__riscv_v)
 #include <riscv_vector.h>
 #endif

@@ -127,6 +129,8 @@ extern float ggml_table_f32_e8m0_half[1 << 8];
 // defined in ggml-cpu.c, initialized in ggml_cpu_init()
 extern float ggml_table_f32_ue4m3[1 << 8];

+extern float ggml_table_f32_f8_e4m3[1 << 8];
+
 // Use lookup table for E8M0 on x86 (faster than bit manipulation)
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 #define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
@@ -1304,7 +1308,7 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
 #define GGML_BF16_FMA_HI(acc, x, y) \
    (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))

-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)

 // compatible with vlen >= 128

@@ -1046,6 +1046,10 @@ class tensor_traits_common : public tensor_traits_base {
                        return true;
                }
            case GGML_OP_FLASH_ATTN_EXT:
+                // K/V row lists are handled by the generic CPU path
+                if (op->src[5] != nullptr) {
+                    return false;
+                }
                forward_flash_attn_ext_f16(params, op);
                return true;
            case GGML_OP_CONT:
@@ -1489,6 +1493,9 @@ static const ggml_backend_buffer_i ggml_backend_riscv64_spacemit_buffer_i = {
    /* .cpy_tensor      = */ nullptr,
    /* .clear           = */ ggml_backend_riscv64_spacemit_buffer_clear,
    /* .reset           = */ nullptr,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };

 static const char * ggml_backend_cpu_riscv64_spacemit_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
@@ -84,7 +84,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
        }
        // reduce sum1,sum2 to sum1
        GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
        int vl = __riscv_vsetvlmax_e32m8();
        vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
        vfloat32m8_t vsum;
@@ -195,7 +195,7 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
    sumf += (ggml_float)_mm_cvtss_f32(g);

 #undef LOAD
-#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfwma)
+#elif defined(__riscv_v) && defined(__riscv_zvfbfwma)
    size_t vl = __riscv_vsetvlmax_e32m4();

    // initialize accumulators to all zeroes
@@ -318,7 +318,7 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
        sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum3_hi);

        sumf = ggml_sve_sum_f32x2(sum1_lo, sum1_hi);
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
        #if defined(__riscv_zvfh)
            int vl = __riscv_vsetvlmax_e32m2();
            vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
@@ -405,7 +405,7 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
    for (; i + 3 < n; i += 4) {
        vec_xst(ggml_v_silu(vec_xl(0, x + i)), 0, y + i);
    }
-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)
    for (int vl; i < n; i += vl) {
        vl = __riscv_vsetvl_e32m2(n - i);
        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
@@ -446,7 +446,7 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
    for (; i + 3 < n; i += 4) {
        vec_xst(vec_mul(ggml_v_silu(vec_xl(0, x + i)), vec_xl(0, g + i)), 0, y + i);
    }
-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)
    for (int vl; i < n; i += vl) {
        vl = __riscv_vsetvl_e32m2(n - i);
        vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
@@ -602,7 +602,7 @@ ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const floa
        val = vec_mul(val, val);
        sum += (ggml_float)vec_hsum_f32x4(val);
    }
-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)
    vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
    for (int vl; i < n; i += vl) {
        vl = __riscv_vsetvl_e32m2(n - i);
@@ -681,7 +681,7 @@ ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float
        vec_xst(v_val, 0, y + i);
        sum += (ggml_float)vec_hsum_f32x4(v_val);
    }
-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)
    vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
    for (int avl; i < n; i += avl) {
        avl = __riscv_vsetvl_e32m2(n - i);
@@ -206,7 +206,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
        sumf[0] = ggml_sve_sum_f32x2(sum_0_lo, sum_0_hi);
        sumf[1] = ggml_sve_sum_f32x2(sum_1_lo, sum_1_hi);
        np = n;
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
        #if defined(__riscv_zvfh)
            size_t vl = __riscv_vsetvlmax_e32m4();

@@ -396,7 +396,7 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const

            svst1_f32(pg, y + np2, ay1);
        }
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
        for (int i = 0, avl; i < n; i += avl) {
            avl = __riscv_vsetvl_e32m8(n - i);
            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
@@ -513,7 +513,7 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
        svst1_f16(pg, (__fp16 *)(y + np2), hy);
    }
    np = n;
-#elif defined(__riscv_v_intrinsic) // implies __riscv_v_intrinsic
+#elif defined(__riscv_v)
    #if defined (__riscv_zvfh)
        const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
        const _Float16 scale = *(const _Float16*)(&s);
@@ -599,7 +599,7 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
                y[i] += x[k][i]*v[k][0];
            }
        }
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
        for (int i = 0, avl; i < n; i += avl) {
            avl = __riscv_vsetvl_e32m8(n - i);
            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
@@ -660,7 +660,7 @@ inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, co
        for (int i = 0; i < n; ++i) {
            y[i] = x[i]*s + b;
        }
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
        for (int i = 0, avl; i < n; i += avl) {
            avl = __riscv_vsetvl_e32m8(n - i);
            vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
@@ -729,7 +729,7 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
            ay1 = svmul_f32_m(pg, ay1, vx);
            svst1_f32(pg, y + i, ay1);
        }
-    #elif defined(__riscv_v_intrinsic)
+    #elif defined(__riscv_v)
        for (int i = 0, avl; i < n; i += avl) {
            avl = __riscv_vsetvl_e32m8(n - i);
            vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
@@ -793,7 +793,7 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
        svst1_f16(pg, (__fp16 *)(y + np), out);
    }
    np = n;
-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)
    #if defined(__riscv_zvfh)
        const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
        const _Float16 scale = *(const _Float16*)(&s);
@@ -1411,7 +1411,7 @@ inline static float32x4_t ggml_v_gelu_quick(float32x4_t x) {
    return vec_div(x, vec_add(one, exp_cx));
 }

-#elif defined(__riscv_v_intrinsic)
+#elif defined(__riscv_v)

 // adapted from arm limited optimized routine
 // the maximum error is 1.45358 plus 0.5 ulps
@@ -1469,7 +1469,7 @@ inline static vfloat32m2_t ggml_v_gelu_quick_m2(vfloat32m2_t x, int 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
+#endif // __ARM_NEON / __AVX2__ / __SSE2__ / __VXE__ / __riscv_v

 inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
    for (int i = 0; i < n; ++i) {
@@ -0,0 +1,198 @@
+#include "batch-get.cuh"
+#include "common.cuh"
+#include "ggml-cpp.h"
+
+#include <climits>
+
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+struct ggml_cuda_batch_get_descriptor {
+    const unsigned char * src;
+    unsigned char * dst;
+    size_t size;
+};
+
+static __global__ void ggml_cuda_batch_get_kernel(const ggml_cuda_batch_get_descriptor * copies, size_t count) {
+    // Blocks along y cooperate on each descriptor.
+    const size_t index = blockIdx.x;
+    if (index >= count) {
+        return;
+    }
+    const size_t first = (size_t) blockIdx.y * blockDim.x + threadIdx.x;
+    const size_t stride = (size_t) gridDim.y * blockDim.x;
+    const auto copy = copies[index];
+    // Use 16-, 8-, 4-, or 2-byte copies when both pointers meet that alignment; threads copy separate chunks.
+    const uintptr_t alignment = (uintptr_t) copy.src | (uintptr_t) copy.dst;
+    size_t tail = 0;
+    if (alignment % 16 == 0) {
+        for (size_t i = first; i < copy.size / 16; i += stride) {
+            ggml_cuda_memcpy_1<16>(copy.dst + 16*i, copy.src + 16*i);
+        }
+        tail = copy.size / 16 * 16;
+    } else if (alignment % 8 == 0) {
+        for (size_t i = first; i < copy.size / 8; i += stride) {
+            ggml_cuda_memcpy_1<8>(copy.dst + 8*i, copy.src + 8*i);
+        }
+        tail = copy.size / 8 * 8;
+    } else if (alignment % 4 == 0) {
+        for (size_t i = first; i < copy.size / 4; i += stride) {
+            ggml_cuda_memcpy_1<4>(copy.dst + 4*i, copy.src + 4*i);
+        }
+        tail = copy.size / 4 * 4;
+    } else if (alignment % 2 == 0) {
+        for (size_t i = first; i < copy.size / 2; i += stride) {
+            ggml_cuda_memcpy_1<2>(copy.dst + 2*i, copy.src + 2*i);
+        }
+        tail = copy.size / 2 * 2;
+    }
+    // Copy trailing bytes, or the entire range when no wider alignment is available.
+    for (size_t i = tail + first; i < copy.size; i += stride) {
+        ggml_cuda_memcpy_1<1>(copy.dst + i, copy.src + i);
+    }
+}
+
+struct ggml_cuda_batch_get_mapping {
+    void * host = nullptr;
+    unsigned char * device = nullptr;
+
+    ggml_cuda_batch_get_mapping(void * host, size_t size) {
+        const cudaError_t err = cudaHostRegister(host, size, cudaHostRegisterMapped | cudaHostRegisterPortable);
+        if (err != cudaSuccess) {
+            (void) cudaGetLastError();
+            GGML_LOG_DEBUG("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
+            return;
+        }
+        this->host = host;
+        if (cudaHostGetDevicePointer((void **) &device, host, 0) != cudaSuccess) {
+            (void) cudaGetLastError();
+            device = nullptr;
+            return;
+        }
+    }
+
+    ggml_cuda_batch_get_mapping(const ggml_cuda_batch_get_mapping &) = delete;
+    ggml_cuda_batch_get_mapping & operator=(const ggml_cuda_batch_get_mapping &) = delete;
+
+    ~ggml_cuda_batch_get_mapping() {
+        if (host) {
+            // Release only the registration acquired by this object.
+            CUDA_CHECK(cudaHostUnregister(host));
+        }
+    }
+};
+#endif
+
+bool ggml_cuda_get_tensor_n(int device,
+        const ggml_tensor * const * tensors, void * const * data, const size_t * offsets,
+        const size_t * sizes, size_t n_tensors) {
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+    GGML_UNUSED(device);
+    GGML_UNUSED(tensors);
+    GGML_UNUSED(data);
+    GGML_UNUSED(offsets);
+    GGML_UNUSED(sizes);
+    GGML_UNUSED(n_tensors);
+    return false;
+#else
+    // Provisional dispatch values based on experiments.
+    constexpr size_t min_batch_reads = 128;
+    constexpr size_t max_gather_read_bytes = 128 * 1024;
+    // Collect non-empty CPU destination ranges.
+    std::vector<std::pair<uintptr_t, uintptr_t>> ranges;
+    ranges.reserve(n_tensors);
+    size_t payload_bytes = 0;
+    // Number of non-empty reads whose size is at most max_gather_read_bytes.
+    size_t n_reads_small = 0;
+    for (size_t i = 0; i < n_tensors; ++i) {
+        if (sizes[i] == 0) {
+            continue;
+        }
+        const uintptr_t address = (uintptr_t) data[i];
+        GGML_ASSERT(address <= UINTPTR_MAX - sizes[i]);
+        GGML_ASSERT(payload_bytes <= SIZE_MAX - sizes[i]);
+        payload_bytes += sizes[i];
+        n_reads_small += sizes[i] <= max_gather_read_bytes;
+        ranges.emplace_back(address, address + sizes[i]);
+    }
+    if (ranges.empty()) {
+        return true;
+    }
+    // Total number of non-empty reads.
+    const size_t n_reads = ranges.size();
+    if (n_reads > SIZE_MAX / sizeof(ggml_cuda_batch_get_descriptor) || n_reads > INT_MAX) {
+        return false;
+    }
+    if (n_reads < 2 || (n_reads < min_batch_reads && n_reads_small == n_reads)) {
+        return false;
+    }
+    std::sort(ranges.begin(), ranges.end());
+    for (size_t i = 1; i < ranges.size(); ++i) {
+        // Preserve the ordered copy path when CPU destinations overlap.
+        if (ranges[i].first < ranges[i - 1].second) {
+            return false;
+        }
+    }
+    // Find the CPU span to register, including gaps between destinations.
+    const uintptr_t begin = ranges.front().first;
+    const uintptr_t end = ranges.back().second;
+    if (end - begin - payload_bytes > payload_bytes) {
+        return false;
+    }
+    // Register and map the existing CPU destination so the kernel can write to it directly.
+    ggml_cuda_batch_get_mapping mapping((void *) begin, end - begin);
+    if (!mapping.host) {
+        return false;
+    }
+    // Reads selected for gather; zero means all reads use DMA.
+    size_t n_reads_gather = mapping.device && n_reads_small >= min_batch_reads ? n_reads_small : 0;
+    // Keep descriptor storage alive until the final stream synchronization.
+    constexpr size_t max_batch_descriptors = 128 * 1024;
+    const size_t bytes = std::min(n_reads_gather, max_batch_descriptors) * sizeof(ggml_cuda_batch_get_descriptor);
+    ggml_backend_buffer_ptr descriptors(n_reads_gather ? ggml_backend_buft_alloc_buffer(ggml_backend_cuda_buffer_type(device), bytes) : nullptr);
+    if (!descriptors) {
+        n_reads_gather = 0;
+    }
+    const auto host_type = ggml_backend_cuda_host_buffer_type();
+    ggml_backend_buffer_ptr host_descriptors(n_reads_gather ? ggml_backend_buft_alloc_buffer(host_type, bytes) : nullptr);
+    if (!host_descriptors || ggml_backend_buffer_get_type(host_descriptors.get()) != host_type) {
+        n_reads_gather = 0;
+    }
+    if (n_reads_gather) {
+        auto * copies = static_cast<ggml_cuda_batch_get_descriptor *>(ggml_backend_buffer_get_base(host_descriptors.get()));
+        auto * device_copies = static_cast<ggml_cuda_batch_get_descriptor *>(ggml_backend_buffer_get_base(descriptors.get()));
+        const int threads = 256;
+        const size_t nsm = ggml_cuda_info().devices[device].nsm;
+        size_t i = 0;
+        for (size_t first = 0; first < n_reads_gather; first += max_batch_descriptors) {
+            const size_t n_batch = std::min(n_reads_gather - first, max_batch_descriptors);
+            size_t count = 0;
+            for (; i < n_tensors && count < n_batch; ++i) {
+                if (sizes[i] == 0 || sizes[i] > max_gather_read_bytes) {
+                    continue;
+                }
+                copies[count++] = {(const unsigned char *) tensors[i]->data + offsets[i],
+                        mapping.device + ((uintptr_t) data[i] - begin), sizes[i]};
+            }
+            const size_t blocks_per_copy = std::min<size_t>(UINT16_MAX, std::max<size_t>(1, (nsm + n_batch - 1) / n_batch));
+            const dim3 blocks((unsigned int) n_batch, (unsigned int) blocks_per_copy);
+            // Stream ordering keeps each kernel ahead of the next descriptor overwrite.
+            CUDA_CHECK(cudaMemcpyAsync(device_copies, copies, n_batch * sizeof(copies[0]), cudaMemcpyHostToDevice, cudaStreamPerThread));
+            ggml_cuda_batch_get_kernel<<<blocks, threads, 0, cudaStreamPerThread>>>(device_copies, n_batch);
+            CUDA_CHECK(cudaGetLastError());
+            if (first + n_batch < n_reads_gather) {
+                // Wait before refilling the pinned host descriptor buffer.
+                CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+            }
+        }
+    }
+    // Keep the destination pinned for DMA too; overlapping destinations already declined above.
+    for (size_t i = 0; i < n_tensors; ++i) {
+        if (sizes[i] && (!n_reads_gather || sizes[i] > max_gather_read_bytes)) {
+            CUDA_CHECK(cudaMemcpyAsync(data[i], (const char *) tensors[i]->data + offsets[i], sizes[i], cudaMemcpyDeviceToHost, cudaStreamPerThread));
+        }
+    }
+    // Wait for GPU writes before releasing the CPU mapping and descriptor buffers.
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+    return true;
+#endif
+}
@@ -0,0 +1,7 @@
+#pragma once
+
+#include "ggml-backend.h"
+
+bool ggml_cuda_get_tensor_n(int device,
+        const ggml_tensor * const * tensors, void * const * data, const size_t * offsets,
+        const size_t * sizes, size_t n_tensors);
@@ -82,6 +82,7 @@
 #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
 #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
 #define GGML_CUDA_CC_RDNA3_5    (GGML_CUDA_CC_OFFSET_AMD + 0x1150) // AI 370, AI Max 395 laptops.
+#define GGML_CUDA_CC_RDNA4_M    (GGML_CUDA_CC_OFFSET_AMD + 0x1170) // gfx117X
 #define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000

 #define GGML_CUDA_CC_IS_AMD(cc)     (cc >= GGML_CUDA_CC_OFFSET_AMD)
@@ -89,8 +90,9 @@
 #define GGML_CUDA_CC_IS_RDNA1(cc)   (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
 #define GGML_CUDA_CC_IS_RDNA2(cc)   (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
 #define GGML_CUDA_CC_IS_RDNA3_0(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA3_5)
-#define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4_M)
 #define GGML_CUDA_CC_IS_RDNA3(cc)   (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc))
+#define GGML_CUDA_CC_IS_RDNA4_M(cc) (cc >= GGML_CUDA_CC_RDNA4_M && cc < GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_RDNA4(cc)   (cc >= GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_GCN_APU(cc) ((cc) == GGML_CUDA_CC_GFX909 || (cc) == GGML_CUDA_CC_GFX90C)
 #define GGML_CUDA_CC_IS_GCN(cc)     ((cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1) || GGML_CUDA_CC_IS_GCN_APU(cc))
@@ -277,9 +279,17 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)

-#if defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA4_M))
+#define AMD_WMMA_128B_AVAILABLE
+#endif // defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA4_M))
+
+#if defined(GGML_USE_HIP) && defined(RDNA3)
+#define AMD_WMMA_256B_AVAILABLE
+#endif // defined(GGML_USE_HIP) && defined(RDNA3)
+
+#if defined(AMD_WMMA_128B_AVAILABLE) || defined(AMD_WMMA_256B_AVAILABLE)
 #define AMD_WMMA_AVAILABLE
-#endif // defined(GGML_USE_HIP) && defined(RDNA4)
+#endif // defined(AMD_WMMA_128B_AVAILABLE) || defined(AMD_WMMA_256B_AVAILABLE)

 // The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
@@ -326,7 +336,7 @@ static bool fast_fp16_hardware_available(const int cc) {
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 }

@@ -355,7 +365,7 @@ static bool amd_mfma_available(const int cc) {
 }

 static bool amd_wmma_available(const int cc) {
-    return (GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_RDNA3(cc));
+    return GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
 }

 static bool volta_mma_available(const int cc) {
@@ -713,7 +723,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #if defined(GGML_USE_HIP)
 #if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
    c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3) || defined(RDNA4)
+#elif defined(RDNA3) || defined(RDNA4_M) || defined(RDNA4)
    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
 #elif defined(RDNA1) || defined(__gfx900__)
    int tmp1;
@@ -758,9 +768,9 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
    acc += v.y*u.y;
 }

-#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4_M) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
 #define V_DOT2_F32_F16_AVAILABLE
-#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4_M) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))

 static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
 #ifdef V_DOT2_F32_F16_AVAILABLE
@@ -1459,16 +1469,21 @@ struct ggml_backend_cuda_context {
    // when the computation is split across CPU/GPU (e.g., with --n-cpu-moe)
    std::unordered_map<const void *, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;

-    int64_t last_graph_eviction_sweep = 0;
+    int64_t next_graph_eviction_time = 0;
+    int64_t last_graph_use_time = 0;

-    ggml_cuda_graph * cuda_graph(const void * first_node_ptr) {
-        const int64_t time_now = ggml_time_us();
+    ggml_cuda_graph * cuda_graph(const void * first_node_ptr, int64_t time_now = ggml_time_us()) {
+        // Give all cached graphs 10s to be reused after an idle period, without updating each graph.
+        if (!cuda_graphs.empty() && time_now - last_graph_use_time >= 10'000'000) {
+            next_graph_eviction_time = time_now + 10'000'000;
+        }
+        last_graph_use_time = time_now;

-        // sweep every 5s, evicting cuda graphs unused for >=10s
-        if (time_now - last_graph_eviction_sweep >= 5'000'000) {
-            last_graph_eviction_sweep = time_now;
+        // Sweep every 5s during active use, evicting graphs unused for at least 10s.
+        if (time_now >= next_graph_eviction_time) {
+            next_graph_eviction_time = time_now + 5'000'000;
            for (auto it = cuda_graphs.begin(); it != cuda_graphs.end(); ) {
-                if (time_now - it->second->last_used_time >= 10'000'000) {
+                if (it->first != first_node_ptr && time_now - it->second->last_used_time >= 10'000'000) {
                    it = cuda_graphs.erase(it);
                } else {
                    ++it;
@@ -1570,6 +1585,7 @@ struct ggml_cuda_mm_fusion_args_host {
    const ggml_tensor * gate_scale = nullptr;
    ggml_glu_op glu_op;
    float glu_limit = 0.0f;
+    ggml_prec prec_src1 = GGML_PREC_UNDEFINED; // src1 precision requested by the fused MUL_MATs (MMQ)
    const ggml_tensor * shared_up = nullptr;
    const ggml_tensor * shared_gate = nullptr;
    ggml_tensor * shared_dst = nullptr;
@@ -25,6 +25,7 @@ typedef void (* fattn_kernel_t)(
        const char * __restrict__ mask,
        const char * __restrict__ sinks,
        const int  * __restrict__ KV_max,
+        const int  * __restrict__ KV_rows,
        float      * __restrict__ dst,
        float2     * __restrict__ dst_meta,
        const float scale,
@@ -427,17 +428,18 @@ static __device__ __forceinline__ void dequantize_V_q4_0(const void * __restrict

    const int8_t * q8 = (const int8_t *) &q;

-#ifdef FP16_AVAILABLE
    if constexpr (std::is_same_v<T, half>) {
+#ifdef FP16_AVAILABLE
        const half2 d = __half2half2(x[ib].d);

 #pragma unroll
        for (int l0 = 0; l0 < ne; l0 += 2) {
            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]);
        }
-    } else
+#else
+        NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
-    if constexpr (std::is_same_v<T, float>) {
+    } else if constexpr (std::is_same_v<T, float>) {
        const float d = x[ib].d;

 #pragma unroll
@@ -465,8 +467,8 @@ static __device__ __forceinline__ void dequantize_V_q4_1(const void * __restrict

    const int8_t * q8 = (const int8_t *) &q;

-#ifdef FP16_AVAILABLE
    if constexpr (std::is_same_v<T, half>) {
+#ifdef FP16_AVAILABLE
        const half2 dm = x[ib].dm;
        const half2 d  = __half2half2( __low2half(dm));
        const half2 m  = __half2half2(__high2half(dm));
@@ -475,9 +477,10 @@ static __device__ __forceinline__ void dequantize_V_q4_1(const void * __restrict
        for (int l0 = 0; l0 < ne; l0 += 2) {
            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]) + m;
        }
-    } else
+#else
+        NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
-    if constexpr (std::is_same_v<T, float>) {
+    } else if constexpr (std::is_same_v<T, float>) {
        const float2 dm = __half22float2(x[ib].dm);

 #pragma unroll
@@ -517,17 +520,18 @@ static __device__ __forceinline__ void dequantize_V_q5_0(const void * __restrict

    const int8_t * q8 = (const int8_t *) &q;

-#ifdef FP16_AVAILABLE
    if constexpr (std::is_same_v<T, half>) {
+#ifdef FP16_AVAILABLE
        const half2 d = __half2half2(x[ib].d);

 #pragma unroll
        for (int l0 = 0; l0 < ne; l0 += 2) {
            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]);
        }
-    } else
+#else
+        NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
-    if constexpr (std::is_same_v<T, float>) {
+    } else if constexpr (std::is_same_v<T, float>) {
        const float d = x[ib].d;

 #pragma unroll
@@ -565,8 +569,8 @@ static __device__ __forceinline__ void dequantize_V_q5_1(const void * __restrict

    const int8_t * q8 = (const int8_t *) &q;

-#ifdef FP16_AVAILABLE
    if constexpr (std::is_same_v<T, half>) {
+#ifdef FP16_AVAILABLE
        const half2 dm = x[ib].dm;
        const half2 d  = __half2half2( __low2half(dm));
        const half2 m  = __half2half2(__high2half(dm));
@@ -575,9 +579,10 @@ static __device__ __forceinline__ void dequantize_V_q5_1(const void * __restrict
        for (int l0 = 0; l0 < ne; l0 += 2) {
            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]) + m;
        }
-    } else
+#else
+        NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
-    if constexpr (std::is_same_v<T, float>) {
+    } else if constexpr (std::is_same_v<T, float>) {
        const float2 dm = __half22float2(x[ib].dm);

 #pragma unroll
@@ -600,17 +605,18 @@ static __device__ __forceinline__ void dequantize_V_q8_0(const void * __restrict
    int8_t qs[ne];
    ggml_cuda_memcpy_1<ne, 2>(qs, x[ib].qs + iqs);

-#ifdef FP16_AVAILABLE
    if constexpr (std::is_same<T, half>::value) {
+#ifdef FP16_AVAILABLE
        const half2 d = __half2half2(x[ib].d);

 #pragma unroll
        for (int l0 = 0; l0 < ne; l0 += 2) {
            ((half2 *) dst)[l0/2] = d * make_half2(qs[l0 + 0], qs[l0 + 1]);
        }
-    } else
+#else
+        NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
-    if constexpr (std::is_same<T, float>::value) {
+    } else if constexpr (std::is_same<T, float>::value) {
        const float d = x[ib].d;

 #pragma unroll
@@ -991,8 +997,13 @@ void launch_fattn(

    const bool V_is_K_view = V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs));

-    const ggml_tensor * mask  = dst->src[3];
-    const ggml_tensor * sinks = dst->src[4];
+    const ggml_tensor * mask    = dst->src[3];
+    const ggml_tensor * sinks   = dst->src[4];
+    const ggml_tensor * kv_rows = dst->src[5];
+
+    // with kv_rows the kernel iterates over the mask columns and reads K/V rows through kv_rows
+    const int64_t n_kv_cols = kv_rows ? kv_rows->ne[0] : K->ne[1];
+    GGML_ASSERT(!kv_rows || !use_sparse);

    ggml_tensor * KQV = dst;

@@ -1113,8 +1124,8 @@ void launch_fattn(

    // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
    // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
-    //     multiple sequences of possibly different lengths.
-    const bool scan_mask = !use_sparse && mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1);
+    //     multiple sequences of possibly different lengths, or kv_rows lists that end with padding.
+    const bool scan_mask = !use_sparse && mask && n_kv_cols % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1 || kv_rows);
    if (scan_mask) {
        const int64_t s31 = mask->nb[1] / sizeof(half2);
        const int64_t s33 = mask->nb[3] / sizeof(half2);
@@ -1123,7 +1134,7 @@ void launch_fattn(
        const dim3 block_dim_KV_max(FATTN_KQ_STRIDE/2, 1, 1);

        const int ne_KV_max = blocks_num_KV_max.x*blocks_num_KV_max.y;
-        const int iter_k = K->ne[1] / FATTN_KQ_STRIDE;
+        const int iter_k = n_kv_cols / FATTN_KQ_STRIDE;

        KV_max.alloc(ne_KV_max);
        ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(blocks_num_KV_max, block_dim_KV_max, 0, main_stream);
@@ -1132,13 +1143,19 @@ void launch_fattn(
        CUDA_CHECK(cudaGetLastError());
    }

+    // K/V are shared by all slices when read through kv_rows
+    if (kv_rows) {
+        nb13 = 0;
+        nb23 = 0;
+    }
+
    const dim3 block_dim(warp_size, nwarps, 1);
    int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy.
    CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared));
    GGML_ASSERT(max_blocks_per_sm > 0);
    int parallel_blocks = max_blocks_per_sm;

-    const int64_t n_kv = use_sparse ? n_kv_max : K->ne[1];
+    const int64_t n_kv = use_sparse ? n_kv_max : n_kv_cols;
    const int ntiles_KV = (n_kv + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by KV cache length.

    dim3 blocks_num;
@@ -1255,6 +1272,7 @@ void launch_fattn(
        mask ? ((const char *) mask->data) : nullptr,
        sinks ? ((const char *) sinks->data) : nullptr,
        KV_max.ptr,
+        kv_rows ? (const int *) kv_rows->data : nullptr,
        !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
        scale, max_bias, m0, m1, n_head_log2, logit_softcap,
        Q->ne[0], ne01,     Q->ne[2], Q->ne[3], Q->nb[1], Q->nb[2], Q->nb[3],
@@ -420,6 +420,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                    // padded slots gather row 0, the -inf mask removes their contribution
                    const int32_t index = i < i_sup ? indices[k_VKQ_0 + i] : 0;
                    i_KV = index >= 0 ? index : 0;
+                } else if (indices) {
+                    // K/V rows given per mask column, negative rows are padding
+                    i_KV = max(indices[k_VKQ_0 + i], 0);
                } else {
                    i_KV = k_VKQ_0 + i;
                }
@@ -467,8 +470,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                    if constexpr (use_sparse) {
                        const int32_t index = i < i_sup ? indices[k_VKQ_0 + i] : -1;
                        src = index >= 0 ? KV + int64_t(index)*stride_KV + k*h2_per_chunk : zero;
+                    } else if (!oob_check || i < i_sup) {
+                        const int64_t i_KV = indices ? max(indices[k_VKQ_0 + i], 0) : k_VKQ_0 + i;
+                        src = KV + i_KV*stride_KV + k*h2_per_chunk;
                    } else {
-                        src = !oob_check || i < i_sup ? KV + int64_t(k_VKQ_0 + i)*stride_KV + k*h2_per_chunk : zero;
+                        src = zero;
                    }
                    ggml_cuda_memcpy_1<16>(swizzle<stride_tile>(tile_KV, i*stride_tile + k*h2_per_chunk, i), src);
                }
@@ -634,7 +640,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        cp_async_wait_all();
        __syncthreads();
        flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
-            (V_h2, tile_V, nbatch_V2, stride_V, k_VKQ_0, k_VKQ_sup, nullptr);
+            (V_h2, tile_V, nbatch_V2, stride_V, k_VKQ_0, k_VKQ_sup, indices);
    } else {
        // the sparse mask values are gathered per element, always load them synchronously
        constexpr bool use_cp_async = nstages == 1 && !use_sparse;
@@ -991,7 +997,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                    (mask_h, tile_mask, stride_mask, k_VKQ_0 + nbatch_fa, k_VKQ_sup, jt*ncols1, ne01, nullptr);
            }
            flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
-                (K_h2, tile_K, nbatch_K2, stride_K, k_VKQ_0 + nbatch_fa, k_VKQ_sup, nullptr);
+                (K_h2, tile_K, nbatch_K2, stride_K, k_VKQ_0 + nbatch_fa, k_VKQ_sup, indices);
        }
    }

@@ -1255,7 +1261,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    T_B_KQ    Q_B[(Q_in_reg ? DKQ/(2*T_B_KQ::J) : 1)];
 #if defined(TURING_MMA_AVAILABLE)
    T_C_VKQ VKQ_C[cols_per_warp == 8 ? DV/T_C_VKQ::I : DV/(2*T_C_VKQ::J)];
-#elif defined(AMD_WMMA_AVAILABLE) && defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
    T_C_VKQ VKQ_C[DV % 32 != 0       ? DV/T_C_VKQ::J : DV/(2*T_C_VKQ::J)];
 #elif defined(AMD_MFMA_AVAILABLE)
    T_C_VKQ VKQ_C[                                     DV/T_C_VKQ::J];
@@ -1343,7 +1349,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                (mask_h, tile_mask, stride_mask, kb0*nbatch_fa, k_VKQ_sup, jt*ncols1, ne01, nullptr);
        }
        flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
-            (K_h2, tile_K, nbatch_K2, stride_K, kb0*nbatch_fa, k_VKQ_sup, nullptr);
+            (K_h2, tile_K, nbatch_K2, stride_K, kb0*nbatch_fa, k_VKQ_sup, indices);
    }

    // kb0_start is always < kb0_stop so the last iter can be executed unconditionally.
@@ -1802,6 +1808,7 @@ static __global__ void flash_attn_ext_f16(
        const char * mask_ptr,
        const char * sinks_ptr,
        const int  * KV_max_ptr,
+        const int  * KV_rows_ptr,
        float      * dst_ptr,
        float2     * dst_meta_ptr,
        const float scale,
@@ -1866,7 +1873,7 @@ static __global__ void flash_attn_ext_f16(
 #endif // defined(AMD_WMMA_AVAILABLE)

 #if defined(AMD_MFMA_AVAILABLE)
-    if (ncols1*ncols2 < 16 || (DKQ > 256 && ncols1*ncols2 < 32)) {
+    if (ncols1*ncols2 < 16 || (DKQ > 256 && ncols1*ncols2 < 32 && !use_sparse)) {
        NO_DEVICE_CODE;
        return;
    }
@@ -1924,7 +1931,8 @@ static __global__ void flash_attn_ext_f16(

        const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
        const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
-        const int32_t * indices = use_sparse ? sparse_indices + (int64_t(sequence % ne33)*iter_j + jt)*ne11 : nullptr;
+        const int32_t * indices = use_sparse ? sparse_indices + (int64_t(sequence % ne33)*iter_j + jt)*ne11 :
+                                  KV_rows_ptr ? KV_rows_ptr + int64_t(sequence)*ne11 : nullptr;

        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;

@@ -1973,7 +1981,8 @@ static __global__ void flash_attn_ext_f16(

    const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
    const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
-    const int32_t * indices = use_sparse ? sparse_indices + (int64_t(sequence % ne33)*iter_j + jt)*ne11 : nullptr;
+    const int32_t * indices = use_sparse ? sparse_indices + (int64_t(sequence % ne33)*iter_j + jt)*ne11 :
+                              KV_rows_ptr ? KV_rows_ptr + int64_t(sequence)*ne11 : nullptr;

    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;

@@ -1989,7 +1998,7 @@ static __global__ void flash_attn_ext_f16(
        (Q_f2, K_h2, V_h2, mask_h, indices, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
         ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
 #else
-    GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
+    GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, KV_rows_ptr, dst_ptr, dst_meta_ptr, scale,
        max_bias, m0, m1, n_head_log2, logit_softcap,
        ne00, ne01, ne02, ne03,
              nb01, nb02, nb03,
@@ -2054,7 +2063,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
    bool use_sparse = false;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_MUSA)
        if constexpr (ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, ncols1, ncols2)) {
            if (ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(cc, dst, ncols1, ncols2)) {
                constexpr bool use_sparse_kernel = true;
@@ -2077,7 +2086,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
                }
            }
        } else
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)
        {
            constexpr bool use_sparse_kernel = false;
            fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view, use_sparse_kernel>;
@@ -797,6 +797,7 @@ static __global__ void flash_attn_tile(
        const char * mask_ptr,
        const char * sinks_ptr,
        const int  * KV_max_ptr,
+        const int  * KV_rows_ptr,
        float      * dst_ptr,
        float2     * dst_meta_ptr,
        const float scale,
@@ -821,6 +822,7 @@ static __global__ void flash_attn_tile(
    const int  * GGML_CUDA_RESTRICT KV_max   = KV_max_ptr;
    float      * GGML_CUDA_RESTRICT dst      = dst_ptr;
    float2     * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
+    GGML_UNUSED(KV_rows_ptr);

    // Skip unused kernel variants for faster compilation:

@@ -1132,7 +1134,7 @@ static __global__ void flash_attn_tile(
        }
    }
 #else
-    GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
+    GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, KV_rows_ptr, dst_ptr, dst_meta_ptr, scale,
        max_bias, m0, m1, n_head_log2, logit_softcap,
        ne00, ne01, ne02, ne03,
              nb01, nb02, nb03,
@@ -25,6 +25,7 @@ static __global__ void flash_attn_ext_vec(
        const char * mask_ptr,
        const char * sinks_ptr,
        const int  * KV_max_ptr,
+        const int  * KV_rows_ptr,
        float      * dst_ptr,
        float2     * dst_meta_ptr,
        const float scale,
@@ -50,6 +51,7 @@ static __global__ void flash_attn_ext_vec(
    const int  * GGML_CUDA_RESTRICT KV_max   = KV_max_ptr;
    float      * GGML_CUDA_RESTRICT dst      = dst_ptr;
    float2     * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
+    GGML_UNUSED(KV_rows_ptr);

    // Skip unused kernel variants for faster compilation:
    if (use_logit_softcap && !(D == 128 || D == 256)) {
@@ -514,7 +516,7 @@ static __global__ void flash_attn_ext_vec(
        dst_meta[((sequence*int(ne01.z) + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
    }
 #else
-    GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
+    GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, KV_rows_ptr, dst_ptr, dst_meta_ptr, scale,
        max_bias, m0, m1, n_head_log2, logit_softcap,
        ne00, ne01, ne02, ne03,
              nb01, nb02, nb03,
@@ -5,7 +5,8 @@
 #include "fattn-vec.cuh"
 #include "fattn.cuh"

-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_MUSA)
+
 // one list per group of ncols1 queries: a column is selected if any query of the group can see it
 template <int ncols1, bool oob>
 __launch_bounds__(256, 1)
@@ -14,10 +15,20 @@ static __global__ void flash_attn_mask_to_sparse_indices(
        const int n_kv_max, const int64_t s31, const int64_t s33) {
    ggml_cuda_pdl_sync();

+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
    constexpr int values_per_lane = 8;
+    using lane_mask_t = std::conditional_t<warp_size == 64, uint64_t, uint32_t>;
+    constexpr uint64_t ballot_mask = ~(lane_mask_t)0;
+    const auto popc = [](const lane_mask_t mask) {
+        if constexpr (warp_size == 64) {
+            return __popcll(mask);
+        } else {
+            return __popc(mask);
+        }
+    };
    const int tid      = threadIdx.x;
-    const int warp     = tid / WARP_SIZE;
-    const int lane     = tid % WARP_SIZE;
+    const int warp     = tid / warp_size;
+    const int lane     = tid % warp_size;
    const int sequence = blockIdx.y;
    const int group    = blockIdx.x;

@@ -27,7 +38,7 @@ static __global__ void flash_attn_mask_to_sparse_indices(
    const half * mask = mask_ptr + sequence*s33 + q0*s31;
    int32_t * indices = indices_ptr + (int64_t(sequence)*gridDim.x + group)*n_kv_max;

-    __shared__ int warp_offsets[256/WARP_SIZE];
+    __shared__ int warp_offsets[256/warp_size];
    __shared__ int row_count;
    __shared__ int chunk_count;

@@ -37,11 +48,11 @@ static __global__ void flash_attn_mask_to_sparse_indices(
    __syncthreads();

    for (int i0 = 0; i0 < ne30; i0 += blockDim.x*values_per_lane) {
-        uint32_t selected_warp[values_per_lane];
+        lane_mask_t selected_warp[values_per_lane];
        int warp_count = 0;
 #pragma unroll
        for (int item = 0; item < values_per_lane; ++item) {
-            const int i = i0 + (warp*values_per_lane + item)*WARP_SIZE + lane;
+            const int i = i0 + (warp*values_per_lane + item)*warp_size + lane;
            bool selected = false;
            if (i < ne30) {
 #pragma unroll
@@ -49,8 +60,8 @@ static __global__ void flash_attn_mask_to_sparse_indices(
                    selected |= (!oob || q < q1 - q0) && isfinite(__half2float(mask[q*s31 + i]));
                }
            }
-            selected_warp[item] = __ballot_sync(0xFFFFFFFF, selected);
-            warp_count += __popc(selected_warp[item]);
+            selected_warp[item] = (lane_mask_t)__ballot_sync(ballot_mask, (int) selected);
+            warp_count += popc(selected_warp[item]);
        }

        if (lane == 0) {
@@ -61,7 +72,7 @@ static __global__ void flash_attn_mask_to_sparse_indices(
        if (tid == 0) {
            int offset = 0;
 #pragma unroll
-            for (int iw = 0; iw < 256/WARP_SIZE; ++iw) {
+            for (int iw = 0; iw < 256/warp_size; ++iw) {
                const int count = warp_offsets[iw];
                warp_offsets[iw] = offset;
                offset += count;
@@ -70,16 +81,16 @@ static __global__ void flash_attn_mask_to_sparse_indices(
        }
        __syncthreads();

-        const uint32_t lane_mask = lane == 0 ? 0 : (1u << lane) - 1;
+        const lane_mask_t lane_mask = lane == 0 ? 0 : ((lane_mask_t)1 << lane) - 1;
        int warp_item_offset = 0;
 #pragma unroll
        for (int item = 0; item < values_per_lane; ++item) {
-            const int i = i0 + (warp*values_per_lane + item)*WARP_SIZE + lane;
-            const int dst = row_count + warp_offsets[warp] + warp_item_offset + __popc(selected_warp[item] & lane_mask);
-            if ((selected_warp[item] & (uint32_t(1) << lane)) && dst < n_kv_max) {
+            const int i = i0 + (warp*values_per_lane + item)*warp_size + lane;
+            const int dst = row_count + warp_offsets[warp] + warp_item_offset + popc(selected_warp[item] & lane_mask);
+            if ((selected_warp[item] & ((lane_mask_t)1 << lane)) && dst < n_kv_max) {
                indices[dst] = i;
            }
-            warp_item_offset += __popc(selected_warp[item]);
+            warp_item_offset += popc(selected_warp[item]);
        }
        __syncthreads();

@@ -101,13 +112,13 @@ static __global__ void flash_attn_mask_to_sparse_indices(
    // the dependent grid reads indices, signal once the row is complete
    ggml_cuda_pdl_lc();
 }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)

 void ggml_cuda_flash_attn_ext_compact_mask(
        const ggml_tensor * mask, int32_t * indices, int32_t * counts, int32_t n_queries, int32_t ncols1, int32_t n_kv_max, cudaStream_t stream) {
-#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+#ifdef GGML_USE_MUSA
    GGML_UNUSED_VARS(mask, indices, counts, n_queries, ncols1, n_kv_max, stream);
-    GGML_ABORT("sparse flash attention is only supported on NVIDIA CUDA");
+    GGML_ABORT("sparse flash attention is not supported on MUSA");
 #else
    const int64_t s31 = mask->nb[1] / sizeof(half);
    const int64_t s33 = mask->nb[3] / sizeof(half);
@@ -122,16 +133,17 @@ void ggml_cuda_flash_attn_ext_compact_mask(
    ggml_cuda_kernel_launch(kernel, launch_params,
        (const half *) mask->data, indices, counts, int(mask->ne[0]), n_queries, n_kv_max, s31, s33);
    CUDA_CHECK(cudaGetLastError());
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)
 }

 bool ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(const int cc, const ggml_tensor * dst, const int ncols1, const int ncols2) {
-#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+#ifdef GGML_USE_MUSA
    GGML_UNUSED_VARS(cc, dst, ncols1, ncols2);
    return false;
 #else
    const ggml_tensor * Q    = dst->src[0];
    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
    const ggml_tensor * mask = dst->src[3];

    float max_bias = 0.0f;
@@ -139,16 +151,21 @@ bool ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(const int cc, const ggml_
    memcpy(&max_bias,      (const float *) dst->op_params + 1, sizeof(float));
    memcpy(&logit_softcap, (const float *) dst->op_params + 2, sizeof(float));

+    // On CDNA only the DKQ=576 kernel variant with ncols2=16 is compiled with sparse support.
+    if (amd_mfma_available(cc) && !ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(int(Q->ne[0]), int(V->ne[0]), 1, 16)) {
+        return false;
+    }
+
    const int32_t n_kv_max = ggml_get_op_params_i32(dst, 4);

    // the dense kernel handles up to 64/ncols2 queries per K/V pass, the single-query gather has to beat that
    const int64_t n_gather = (ncols1 == 1 ? std::min<int64_t>(Q->ne[1], 64/ncols2) : ncols1) * (int64_t) n_kv_max;

-    return GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc) &&
+    return ((GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc)) || amd_mfma_available(cc)) && dst->src[5] == nullptr &&
        mask != nullptr && n_kv_max > 0 && max_bias == 0.0f && logit_softcap == 0.0f &&
        mask->ne[0] == K->ne[1] && mask->ne[1] >= Q->ne[1] && mask->ne[2] == 1 &&
-        K->ne[1] >= std::max<int64_t>(4096, 2*n_gather);
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+        K->ne[1] >= std::max<int64_t>(4096, GGML_CUDA_CC_IS_NVIDIA(cc) ? 2*n_gather : 2LL*n_kv_max);
+#endif // !defined(GGML_USE_MUSA)
 }

 template <int DKQ, int DV, int ncols2>
@@ -156,7 +173,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
    const ggml_tensor * Q = dst->src[0];

-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_MUSA)
    if constexpr (ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, 1, ncols2)) {
        // a sparse variant at the full tile width gathers the union of its queries once, prefer it for large batches
        constexpr bool has_wide_sparse = ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, 64/ncols2, ncols2);
@@ -165,7 +182,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
            return;
        }
    }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)

    if constexpr (ncols2 <= 8) {
        if (turing_mma_available(cc) && Q->ne[1] <= 8/ncols2) {
@@ -199,12 +216,15 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
    const ggml_tensor * V    = dst->src[2];
    const ggml_tensor * mask = dst->src[3];

+    // with kv_rows the kernel iterates over the mask columns
+    const int64_t n_kv = dst->src[5] ? dst->src[5]->ne[0] : K->ne[1];
+
    float max_bias = 0.0f;
    memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));

    // Edge cases like no mask, ALiBi, unpadded K/V, or misaligned addresses for large data transfers
    //     are put into the template specialization without GQA optimizations.
-    bool use_gqa_opt = mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    bool use_gqa_opt = mask && max_bias == 0.0f && n_kv % FATTN_KQ_STRIDE == 0;
    for (const ggml_tensor * t : {Q, K, V, mask}) {
        if (t == nullptr || ggml_is_quantized(t->type)) {
            continue;
@@ -365,7 +385,9 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg

            GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
            const int gqa_ratio = Q->ne[2] / K->ne[2];
-            if (gqa_ratio == 20 && GGML_CUDA_CC_IS_NVIDIA(cc)) { // GLM 4.7 Flash
+            if (amd_mfma_available(cc) && ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(cc, dst, 1, 16)) {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            } else if (gqa_ratio == 20 && GGML_CUDA_CC_IS_NVIDIA(cc)) { // GLM 4.7 Flash
                if (cc >= GGML_CUDA_CC_DGX_SPARK) {
                    if (Q->ne[1] <= 8) {
                        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
@@ -548,6 +570,10 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
    const ggml_tensor * K     = dst->src[1];
    const ggml_tensor * V     = dst->src[2];
    const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * kv_rows = dst->src[5];
+
+    // with kv_rows the kernel iterates over the mask columns
+    const int64_t n_kv = kv_rows ? kv_rows->ne[0] : K->ne[1];

    const int gqa_ratio = Q->ne[2] / K->ne[2];
    GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
@@ -557,7 +583,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const

    // The effective batch size for the kernel can be increased by gqa_ratio.
    // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
-    bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && n_kv % FATTN_KQ_STRIDE == 0;
    for (const ggml_tensor * t : {Q, K, V, mask}) {
        if (t == nullptr || ggml_is_quantized(t->type)) {
            continue;
@@ -629,6 +655,12 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
        return BEST_FATTN_KERNEL_NONE;
    }

+    // only the MMA kernel reads K/V through kv_rows
+    if (kv_rows) {
+        return GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc) && mask && Q->ne[0] != 40 && Q->ne[0] != 72 ?
+            BEST_FATTN_KERNEL_MMA_F16 : BEST_FATTN_KERNEL_NONE;
+    }
+
    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
    // 192 satisfies % 64 == 0 but has no vec instance (DKQ != DV); force it onto the MMA path.
    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && Q->ne[0] != 192 && K->ne[1] % FATTN_KQ_STRIDE == 0;
@@ -679,6 +711,11 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
        return BEST_FATTN_KERNEL_MMA_F16;
    }

+    // On CDNA, the sparse MMA kernel is used for DKQ=576 regardless of the head size heuristics.
+    if (amd_mfma_available(cc) && ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(cc, dst, 1, 16)) {
+        return BEST_FATTN_KERNEL_MMA_F16;
+    }
+
    // AMD MFMA needs a certain minimum batch size to outscale the tile kernel for large head sizes.
    if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
        if ((Q->ne[0] <= 64 && Q->ne[1] * gqa_ratio_eff > 8)) {
@@ -9,6 +9,7 @@
 #include "ggml-cuda/arange.cuh"
 #include "ggml-cuda/argmax.cuh"
 #include "ggml-cuda/argsort.cuh"
+#include "ggml-cuda/batch-get.cuh"
 #include "ggml-cuda/binbcast.cuh"
 #include "ggml-cuda/clamp.cuh"
 #include "ggml-cuda/col2im-1d.cuh"
@@ -861,6 +862,51 @@ static void ggml_backend_cuda_buffer_get_tensor_2d(ggml_backend_buffer_t buffer,
    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }

+static void ggml_backend_cuda_buffer_set_tensor_n(ggml_backend_buffer_t buffer, struct ggml_tensor * const * tensors, const void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        if (sizes[i] == 0) {
+            continue;
+        }
+        CUDA_CHECK(cudaMemcpyAsync(
+            (char *) tensors[i]->data + offsets[i],
+            data[i],
+            sizes[i],
+            cudaMemcpyHostToDevice,
+            cudaStreamPerThread));
+    }
+
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
+static void ggml_backend_cuda_buffer_get_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * tensors, void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+
+    if (!ggml_cuda_pinned_memory_disabled() &&
+            ggml_cuda_get_tensor_n(ctx->device, tensors, data, offsets, sizes, n_tensors)) {
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        if (sizes[i] == 0) {
+            continue;
+        }
+        CUDA_CHECK(cudaMemcpyAsync(
+            data[i],
+            (const char *) tensors[i]->data + offsets[i],
+            sizes[i],
+            cudaMemcpyDeviceToHost,
+            cudaStreamPerThread));
+    }
+
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
 static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
    if (ggml_backend_buffer_is_cuda(src->buffer)) {
        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
@@ -886,6 +932,54 @@ static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, co
    GGML_UNUSED(buffer);
 }

+static bool ggml_backend_cuda_buffer_cpy_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors) {
+    ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+        if (!ggml_backend_buffer_is_cuda(src_buf)) {
+            return false;
+        }
+    }
+
+#ifdef GGML_CUDA_NO_PEER_COPY
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+        ggml_backend_buffer_t dst_buf = dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src_buf->context;
+        ggml_backend_cuda_buffer_context * dst_ctx_i = (ggml_backend_cuda_buffer_context *)dst_buf->context;
+        const int src_physical = ggml_cuda_get_physical_device(src_ctx->device);
+        const int dst_physical = ggml_cuda_get_physical_device(dst_ctx_i->device);
+        if (src_physical != dst_physical) {
+            return false;
+        }
+    }
+#endif
+
+    ggml_cuda_set_device(dst_ctx->device);
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+        ggml_backend_buffer_t dst_buf = dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src_buf->context;
+        ggml_backend_cuda_buffer_context * dst_ctx_i = (ggml_backend_cuda_buffer_context *)dst_buf->context;
+        const int src_physical = ggml_cuda_get_physical_device(src_ctx->device);
+        const int dst_physical = ggml_cuda_get_physical_device(dst_ctx_i->device);
+        if (src_physical == dst_physical) {
+            CUDA_CHECK(cudaMemcpyAsync(dsts[i]->data, srcs[i]->data, ggml_nbytes(srcs[i]), cudaMemcpyDeviceToDevice, cudaStreamPerThread));
+        } else {
+#ifdef GGML_CUDA_NO_PEER_COPY
+            GGML_ABORT("peer copy not supported");
+#else
+            CUDA_CHECK(cudaMemcpyPeerAsync(dsts[i]->data, dst_physical, srcs[i]->data, src_physical, ggml_nbytes(srcs[i]), cudaStreamPerThread));
+#endif
+        }
+    }
+
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+    return true;
+}
+
 static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;

@@ -906,6 +1000,9 @@ static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
    /* .cpy_tensor      = */ ggml_backend_cuda_buffer_cpy_tensor,
    /* .clear           = */ ggml_backend_cuda_buffer_clear,
    /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ ggml_backend_cuda_buffer_set_tensor_n,
+    /* .get_tensor_n    = */ ggml_backend_cuda_buffer_get_tensor_n,
+    /* .cpy_tensor_n    = */ ggml_backend_cuda_buffer_cpy_tensor_n,
 };

 // cuda buffer type
@@ -1026,6 +1123,18 @@ struct ggml_backend_cuda_comm_context {

    ggml_cuda_ar_pipeline *     ar_pipeline = nullptr;

+    // Number of ranks in the communicator when each rank is a separate process. Zero unless
+    // comm_init_rank created this context.
+    int                         world_size = 0;
+
+    // Number of ranks taking part in the reduction. For a communicator from comm_init_rank this is
+    // world_size. Otherwise it is the number of backends this process holds. The NCCL reduction
+    // uses it to choose the tensor size below which it reduces in F32 rather than converting to
+    // BF16 first.
+    size_t n_ranks() const {
+        return world_size > 0 ? (size_t) world_size : backends.size();
+    }
+
 #ifdef GGML_USE_NCCL
    std::vector<ncclComm_t>     comms;
 #endif // GGML_USE_NCCL
@@ -1053,6 +1162,7 @@ static bool ggml_backend_cuda_comm_allreduce_nccl(
    }

    const size_t n_backends = comm_ctx->backends.size();
+    const size_t n_ranks    = comm_ctx->n_ranks();

    for (size_t i = 0; i < n_backends; ++i) {
        GGML_ASSERT(tensors[i] != nullptr);
@@ -1062,7 +1172,7 @@ static bool ggml_backend_cuda_comm_allreduce_nccl(

    // For small tensors, simply reduce them as FP32.
    // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
-    if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
+    if ((n_ranks <= 2 && ne < 32768) || (n_ranks == 3 && ne < 131072) || (n_ranks >= 4 && ne < 262144)) {
        for (size_t i = 0; i < n_backends; ++i) {
            if ((tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
                ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
@@ -1293,6 +1403,90 @@ static void * ggml_backend_cuda_comm_init(ggml_backend_t * backends, size_t n_ba
    return ret;
 }

+#ifdef GGML_USE_NCCL
+// Builds the communicator with ncclCommInitRank, for the case where each rank is its own
+// process. Blocks until all world_size ranks have joined.
+static void * ggml_backend_cuda_comm_init_rank(ggml_backend_t backend, int rank, int world_size,
+                                               ggml_backend_comm_bootstrap_callback bootstrap_callback, void * bootstrap_callback_data) {
+    if (world_size < 1 || rank < 0 || rank >= world_size) {
+        GGML_LOG_ERROR("%s: invalid rank %d for world size %d\n", __func__, rank, world_size);
+        return nullptr;
+    }
+
+    // Every rank sends whether its local checks passed, and rank 0 also sends the NCCL unique id.
+    // So all ranks either join together or fail together, and no rank waits in ncclCommInitRank alone.
+    struct bootstrap_msg {
+        uint8_t      ok;
+        ncclUniqueId id;
+    };
+    bootstrap_msg local;
+    memset(&local, 0, sizeof(local));
+    local.ok = 1;
+
+    if (!ggml_backend_is_cuda(backend)) {
+        GGML_LOG_ERROR("%s: backend is not a CUDA backend\n", __func__);
+        local.ok = 0;
+    } else {
+        // NCCL requires one distinct physical GPU per rank, see ggml_backend_cuda_comm_init_nccl.
+        const ggml_cuda_device_info & info = ggml_cuda_info();
+        if (info.device_count > info.physical_device_count) {
+            GGML_LOG_ERROR("%s: CUDA virtual devices are in use, which NCCL does not support\n", __func__);
+            local.ok = 0;
+        }
+    }
+    if (local.ok && rank == 0) {
+        const ncclResult_t rc = ncclGetUniqueId(&local.id);
+        if (rc != ncclSuccess) {
+            GGML_LOG_ERROR("%s: ncclGetUniqueId failed (%s)\n", __func__, ncclGetErrorString(rc));
+            local.ok = 0;
+        }
+    }
+
+    std::vector<bootstrap_msg> all(world_size);
+    if (!bootstrap_callback(bootstrap_callback_data, &local, all.data(), sizeof(local))) {
+        GGML_LOG_ERROR("%s: bootstrap exchange failed on rank %d\n", __func__, rank);
+        return nullptr;
+    }
+    for (int i = 0; i < world_size; i++) {
+        if (!all[i].ok) {
+            GGML_LOG_ERROR("%s: rank %d failed its setup checks\n", __func__, i);
+            return nullptr;
+        }
+    }
+
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
+
+    auto * ret = new ggml_backend_cuda_comm_context;
+    ret->backends.push_back(backend);
+    ret->dev_ids.push_back(cuda_ctx->device);
+    ret->world_size = world_size;
+    ret->comms.resize(1);
+    ret->try_allreduce = ggml_backend_cuda_comm_try_allreduce_nccl;
+
+    // On DGX Spark, NCCL picks 64 channels by default, which is slower and uses more memory than 8.
+    // NCCL_MAX_CTAS still overrides this.
+    ncclConfig_t config = NCCL_CONFIG_INITIALIZER;
+    const int cc = ggml_cuda_info().devices[cuda_ctx->device].cc;
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_DGX_SPARK) {
+        config.maxCTAs = 8;
+    }
+
+    ggml_cuda_set_device(cuda_ctx->device);
+    const ncclResult_t rc = ncclCommInitRankConfig(&ret->comms[0], world_size, all[0].id, rank, &config);
+    if (rc != ncclSuccess) {
+        GGML_LOG_ERROR("%s: ncclCommInitRankConfig failed for rank %d/%d (%s)\n",
+                       __func__, rank, world_size, ncclGetErrorString(rc));
+        ret->comms.clear();
+        delete ret;
+        return nullptr;
+    }
+
+    GGML_LOG_INFO("%s: joined NCCL communicator as rank %d/%d on device %d\n",
+                  __func__, rank, world_size, cuda_ctx->device);
+    return ret;
+}
+#endif // GGML_USE_NCCL
+
 // Top-level dispatch -- calls the function pointer chosen by comm_init.
 // Returns false to let the meta-backend's butterfly run.
 static bool ggml_backend_cuda_comm_allreduce_tensor(void * comm_ctx_v, struct ggml_tensor ** tensors) {
@@ -1565,7 +1759,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
    if (compute_type == GGML_TYPE_F16) {
        prefer_f32_output = cc == GGML_CUDA_CC_VOLTA || GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_CDNA(cc);
    } else if (compute_type == GGML_TYPE_BF16) {
-        prefer_f32_output = !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc);
+        prefer_f32_output = !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_RDNA4_M(cc) && !GGML_CUDA_CC_IS_CDNA(cc);
    }

    if (prefer_f32_output) {
@@ -1957,8 +2151,50 @@ static bool ggml_cuda_match_shared_expert(const ggml_cgraph * graph, int routed_
    return true;
 }

+// The NVFP4 weight scale that scale_node (MUL) applies to the output of mm_node (MUL_MAT), or nullptr.
+// Shared by ggml_cuda_try_fuse and graph_optimize so that both accept the same patterns.
+static const ggml_tensor * ggml_cuda_get_mul_mat_scale(const ggml_tensor * scale_node, const ggml_tensor * mm_node) {
+    const bool scale_lhs_mm = scale_node->src[0] == mm_node;
+    const bool scale_rhs_mm = scale_node->src[1] == mm_node;
+    if (!scale_lhs_mm && !scale_rhs_mm) {
+        return nullptr;
+    }
+
+    const ggml_tensor * scale = scale_lhs_mm ? scale_node->src[1] : scale_node->src[0];
+    if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
+            scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != 1 ||
+            !ggml_are_same_shape(scale_node, mm_node)) {
+        return nullptr;
+    }
+
+    return scale;
+}
+
+// The same for MUL_MAT_ID with one scale per expert: RESHAPE, REPEAT and GET_ROWS (by the ids) of the scales, then the MUL.
+static const ggml_tensor * ggml_cuda_get_mul_mat_id_scale(const ggml_tensor * reshape, const ggml_tensor * repeat, const ggml_tensor * getrows,
+        const ggml_tensor * scale_node, const ggml_tensor * mm_node) {
+    if (repeat->src[0] != reshape || getrows->src[0] != repeat || getrows->src[1] != mm_node->src[2]) {
+        return nullptr;
+    }
+    if (!((scale_node->src[0] == mm_node && scale_node->src[1] == getrows) ||
+            (scale_node->src[0] == getrows && scale_node->src[1] == mm_node))) {
+        return nullptr;
+    }
+
+    const ggml_tensor * scale = reshape->src[0];
+    if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
+            scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != mm_node->src[0]->ne[2] ||
+            !ggml_are_same_shape(scale_node, mm_node)) {
+        return nullptr;
+    }
+
+    return scale;
+}
+
 // 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) {
+// up_scale/gate_scale: the MUL nodes that apply the NVFP4 weight scales (per tensor or per expert) between the MUL_MATs and the GLU.
+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 * up_scale = nullptr, const ggml_tensor * gate_scale = nullptr) {
    const ggml_tensor * x_up = up->src[0];
    const ggml_tensor * y    = up->src[1];

@@ -1971,7 +2207,12 @@ static bool ggml_cuda_mul_mat_q_fusion_matches(const ggml_tensor * up, const ggm
        return false;
    }

-    if (!ggml_cuda_should_fuse_mul_mat(up, gate, glu)) {
+    if (!ggml_cuda_should_fuse_mul_mat(up, gate, glu, nullptr, nullptr, up_scale, gate_scale)) {
+        return false;
+    }
+
+    // The fused kernel runs both MUL_MATs at one src1 precision, see ggml_cuda_mul_mat_q.
+    if (ggml_prec_get_src(up, 1) != ggml_prec_get_src(gate, 1)) {
        return false;
    }

@@ -1984,7 +2225,8 @@ static bool ggml_cuda_mul_mat_q_fusion_matches(const ggml_tensor * up, const ggm
    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) {
+static bool ggml_cuda_should_fuse_mul_mat_q(const ggml_tensor * up, const ggml_tensor * gate, const ggml_tensor * glu,
+        const ggml_tensor * up_scale = nullptr, const ggml_tensor * gate_scale = nullptr) {
    // 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;
@@ -1996,7 +2238,7 @@ static bool ggml_cuda_should_fuse_mul_mat_q(const ggml_tensor * up, const ggml_t
            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])) {
+    if (!ggml_cuda_mul_mat_q_fusion_matches(up, gate, glu, device, up_scale, gate_scale) || bad_padding_clear(up->src[0]) || bad_padding_clear(gate->src[0])) {
        return false;
    }

@@ -2015,7 +2257,7 @@ static bool ggml_cuda_should_fuse_mul_mat_q(const ggml_tensor * up, const ggml_t
    }

    // 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], ncols, nchannels, device);
+    return ggml_cuda_mmq_fusion_is_efficient(x_up->type, ggml_prec_get_src(up, 1), x_up->ne[1], ncols, nchannels, device);
 }

 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -4090,43 +4332,6 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
    bool fused_mul_mat_vec = false;
    int  fused_node_count  = 0;

-    auto get_mul_mat_scale = [](const ggml_tensor * scale_node, const ggml_tensor * mm_node) -> const ggml_tensor * {
-        const bool scale_lhs_mm = scale_node->src[0] == mm_node;
-        const bool scale_rhs_mm = scale_node->src[1] == mm_node;
-        if (!scale_lhs_mm && !scale_rhs_mm) {
-            return nullptr;
-        }
-
-        const ggml_tensor * scale = scale_lhs_mm ? scale_node->src[1] : scale_node->src[0];
-        if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
-                scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != 1 ||
-                !ggml_are_same_shape(scale_node, mm_node)) {
-            return nullptr;
-        }
-
-        return scale;
-    };
-
-    auto get_mul_mat_id_scale = [](const ggml_tensor * reshape, const ggml_tensor * repeat, const ggml_tensor * getrows,
-            const ggml_tensor * scale_node, const ggml_tensor * mm_node) -> const ggml_tensor * {
-        if (repeat->src[0] != reshape || getrows->src[0] != repeat || getrows->src[1] != mm_node->src[2]) {
-            return nullptr;
-        }
-        if (!((scale_node->src[0] == mm_node && scale_node->src[1] == getrows) ||
-                (scale_node->src[0] == getrows && scale_node->src[1] == mm_node))) {
-            return nullptr;
-        }
-
-        const ggml_tensor * scale = reshape->src[0];
-        if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
-                scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != mm_node->src[0]->ne[2] ||
-                !ggml_are_same_shape(scale_node, mm_node)) {
-            return nullptr;
-        }
-
-        return scale;
-    };
-
    auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) -> const ggml_tensor * {
        if (op_bias == GGML_OP_ADD) {
            if (bias_node->src[0] == mul_node) {
@@ -4193,8 +4398,8 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                    continue;
                }

-                const ggml_tensor * gate_scale = get_mul_mat_scale(gate_scale_n, gate_n);
-                const ggml_tensor * up_scale   = get_mul_mat_scale(up_scale_n, up_n);
+                const ggml_tensor * gate_scale = ggml_cuda_get_mul_mat_scale(gate_scale_n, gate_n);
+                const ggml_tensor * up_scale   = ggml_cuda_get_mul_mat_scale(up_scale_n, up_n);
                if (!gate_scale || !up_scale) {
                    continue;
                }
@@ -4225,6 +4430,14 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                    fused_node_count  = n_ops;
                    break;
                }
+
+                if (!with_bias && ggml_cuda_should_fuse_mul_mat_q(up_n, gate_n, glu, up_scale_n, gate_scale_n)) {
+                    fusion_data.prec_src1 = ggml_prec_get_src(up_n, 1);
+                    ggml_cuda_mul_mat_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
+                    fused_mul_mat_vec = true;
+                    fused_node_count  = n_ops;
+                    break;
+                }
            }

            if (fused_mul_mat_vec) {
@@ -4289,9 +4502,9 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                    continue;
                }

-                const ggml_tensor * gate_scale = get_mul_mat_id_scale(cgraph->nodes[gate_idx + 1], cgraph->nodes[gate_idx + 2],
+                const ggml_tensor * gate_scale = ggml_cuda_get_mul_mat_id_scale(cgraph->nodes[gate_idx + 1], cgraph->nodes[gate_idx + 2],
                        cgraph->nodes[gate_idx + 3], gate_scale_n, gate_n);
-                const ggml_tensor * up_scale = get_mul_mat_id_scale(cgraph->nodes[up_idx + 1], cgraph->nodes[up_idx + 2],
+                const ggml_tensor * up_scale = ggml_cuda_get_mul_mat_id_scale(cgraph->nodes[up_idx + 1], cgraph->nodes[up_idx + 2],
                        cgraph->nodes[up_idx + 3], up_scale_n, up_n);
                if (!gate_scale || !up_scale) {
                    continue;
@@ -4319,6 +4532,14 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                    fused_node_count  = n_ops;
                    break;
                }
+
+                if (!with_bias && ggml_cuda_should_fuse_mul_mat_q(up_n, gate_n, glu, up_scale_n, gate_scale_n)) {
+                    fusion_data.prec_src1 = ggml_prec_get_src(up_n, 1);
+                    ggml_cuda_mul_mat_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
+                    fused_mul_mat_vec = true;
+                    fused_node_count  = n_ops;
+                    break;
+                }
            }

            if (fused_mul_mat_vec) {
@@ -4434,6 +4655,7 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                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);
+                fusion_data.prec_src1 = ggml_prec_get_src(up, 1);

                ggml_cuda_mul_mat_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
                fused_mul_mat_vec = true;
@@ -4495,9 +4717,9 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph

            const ggml_tensor * scale = nullptr;
            if (op == GGML_OP_MUL_MAT) {
-                scale = get_mul_mat_scale(scale_node, mm_node);
+                scale = ggml_cuda_get_mul_mat_scale(scale_node, mm_node);
            } else {
-                scale = get_mul_mat_id_scale(cgraph->nodes[i + 1], cgraph->nodes[i + 2], cgraph->nodes[i + 3], scale_node, mm_node);
+                scale = ggml_cuda_get_mul_mat_id_scale(cgraph->nodes[i + 1], cgraph->nodes[i + 2], cgraph->nodes[i + 3], scale_node, mm_node);
            }
            if (!scale) {
                continue;
@@ -5048,6 +5270,33 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
                    continue;
                }
            }
+            // Same with the NVFP4 scales: MUL_MAT(gate), MUL(gate scale), MUL_MAT(up), MUL(up scale), GLU.
+            if (i + 4 < cgraph->n_nodes && cgraph->nodes[i + 4]->op == GGML_OP_GLU) {
+                ggml_tensor * glu        = cgraph->nodes[i + 4];
+                ggml_tensor * gate_scale = cgraph->nodes[i + 1];
+                ggml_tensor * up_scale   = cgraph->nodes[i + 3];
+                if (glu->src[0] == gate_scale && glu->src[1] == up_scale &&
+                        ggml_cuda_mul_mat_q_fusion_matches(cgraph->nodes[i + 2], cgraph->nodes[i], glu, cuda_ctx->device, up_scale, gate_scale) &&
+                        ggml_cuda_get_mul_mat_scale(gate_scale, cgraph->nodes[i]) && ggml_cuda_get_mul_mat_scale(up_scale, cgraph->nodes[i + 2])) {
+                    add_alloc_deps(i, i + 4);
+                    i += 4;
+                    continue;
+                }
+            }
+            // Same with per-expert NVFP4 scales: MUL_MAT_ID, RESHAPE, REPEAT, GET_ROWS, MUL for gate and for up, GLU.
+            if (i + 10 < cgraph->n_nodes && cgraph->nodes[i + 10]->op == GGML_OP_GLU) {
+                ggml_tensor * glu        = cgraph->nodes[i + 10];
+                ggml_tensor * gate_scale = cgraph->nodes[i + 4];
+                ggml_tensor * up_scale   = cgraph->nodes[i + 9];
+                if (glu->src[0] == gate_scale && glu->src[1] == up_scale &&
+                        ggml_cuda_mul_mat_q_fusion_matches(cgraph->nodes[i + 5], cgraph->nodes[i], glu, cuda_ctx->device, up_scale, gate_scale) &&
+                        ggml_cuda_get_mul_mat_id_scale(cgraph->nodes[i + 1], cgraph->nodes[i + 2], cgraph->nodes[i + 3], gate_scale, cgraph->nodes[i]) &&
+                        ggml_cuda_get_mul_mat_id_scale(cgraph->nodes[i + 6], cgraph->nodes[i + 7], cgraph->nodes[i + 8], up_scale, cgraph->nodes[i + 5])) {
+                    add_alloc_deps(i, i + 10);
+                    i += 10;
+                    continue;
+                }
+            }

            ggml_cuda_moe_weighted_reduction_match match;
            if (ggml_cuda_match_moe_weighted_reduction(cgraph, i, match)) {
@@ -6293,6 +6542,11 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con
    if (strcmp(name, "ggml_backend_comm_allreduce_tensor") == 0) {
        return (void *)ggml_backend_cuda_comm_allreduce_tensor;
    }
+#ifdef GGML_USE_NCCL
+    if (strcmp(name, "ggml_backend_comm_init_rank") == 0) {
+        return (void *)ggml_backend_cuda_comm_init_rank;
+    }
+#endif // GGML_USE_NCCL
    if (strcmp(name, "ggml_backend_register_host_buffer") == 0) {
        return (void *)ggml_backend_cuda_register_host_buffer;
    }
@@ -2,239 +2,7 @@
 #include "lightning-indexer.cuh"
 #include "fattn-common.cuh"
 #include "convert.cuh"
-
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-#if defined(TURING_MMA_AVAILABLE)
-
-typedef union {
-    int2 i2;
-    half2 h2[2];
-} half4;
-
-// TODO add support for AMD cards via rocWMMA
-#include <mma.h>
-namespace wmma = nvcuda::wmma;
-
-template <int WARPS_PER_BLOCK, int K_VECS_PER_BLOCK, int64_t N_EMBD, int64_t N_HEAD, ggml_type TYPE_K>
-static __global__ void lightning_indexer_kernel_wmma(
-        const float * Q, const char * K, const float * W, const half * M, float * dst,
-        int64_t n_stream, int64_t n_batch, int64_t n_kv,
-        size_t nb1, size_t nb2, size_t nb3,
-        size_t nbq1, size_t nbq2, size_t nbq3,
-        size_t nbk1, size_t nbk2, size_t nbk3,
-        size_t nbw1, size_t nbw2, size_t nbw3,
-        size_t nbm1, size_t nbm2, size_t nbm3,
-        int64_t nem3
-    ) {
-
-    constexpr int THREADS_PER_BLOCK = WARPS_PER_BLOCK * WARP_SIZE;
-    constexpr int HEADS_PER_INNER_LOOP = 8;
-    constexpr int K_EMBD_PER_INNER_LOOP = 16;
-    constexpr int N_EMBD_PADDED = N_EMBD + 8;
-
-    const int i_batch  = blockIdx.y;
-    const int i_stream = blockIdx.z;
-    const int i_warp   = threadIdx.y;
-    const int i_lane   = threadIdx.x;
-    const int tid      = i_warp * WARP_SIZE + i_lane;
-
-    // each block processes K_VECS_PER_BLOCK K vectors
-    const int start_kv = blockIdx.x * K_VECS_PER_BLOCK;
-
-    const char  * q_base = (const char  *)                 Q + i_batch*nbq2 + i_stream*nbq3;
-    const float * w_base = (const float *) ((const char *) W + i_batch*nbw1 + i_stream*nbw3);
-
-    // phase 1 - load weights and first Q tile to shared memory
-
-    __shared__ float w_shared[N_HEAD];
-    __shared__ int2  q_shared_h[HEADS_PER_INNER_LOOP][N_EMBD_PADDED / 4];
-
-    if (tid < N_HEAD) {
-        w_shared[tid] = w_base[tid];
-    }
-
-    // total number of half4 elements in HEADS_PER_INNER_LOOP x N_EMBD Q tile
-    constexpr int N_Q_TILE = HEADS_PER_INNER_LOOP * (N_EMBD / 4);
-    // number of registers needed in each thread to store Q tile in thread block
-    constexpr int N_Q_NEXT = (N_Q_TILE + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
-
-#pragma unroll
-    for (int i_q = tid; i_q < N_Q_TILE; i_q += THREADS_PER_BLOCK) {
-        const int i_head = i_q / (N_EMBD / 4);
-        const int i_embd = i_q % (N_EMBD / 4);
-        const float4 q = *(const float4 *) (q_base + i_head*nbq1 + i_embd*sizeof(float4));
-        half4 q_packed;
-        q_packed.h2[0] = __float22half2_rn(make_float2(q.x, q.y));
-        q_packed.h2[1] = __float22half2_rn(make_float2(q.z, q.w));
-        q_shared_h[i_head][i_embd] = q_packed.i2;
-    }
-
-    // phase 2 - load (and dequantize if needed) K to shared mem
-
-    __shared__ half2 k_shared_h[K_VECS_PER_BLOCK][N_EMBD_PADDED / 4][2];
-
-    constexpr int n_k = K_VECS_PER_BLOCK * (N_EMBD / 4);
-
-    if constexpr (TYPE_K == GGML_TYPE_F16) {
-#pragma unroll
-        for (int i_k = tid; i_k < n_k; i_k += THREADS_PER_BLOCK) {
-            const int i_k_vec = i_k / (N_EMBD / 4);
-            const int i_embd = i_k % (N_EMBD / 4);
-            const int i_kv = start_kv + i_k_vec;
-            if (i_kv < n_kv) {
-                const int2 * k_base = (const int2 *) ((const char *) K + i_kv*nbk2 + i_stream*nbk3);
-                *(int2*) &k_shared_h[i_k_vec][i_embd] = k_base[i_embd];
-            } else {
-                *(int2*) &k_shared_h[i_k_vec][i_embd] = make_int2(0, 0);
-            }
-        }
-    } else {
-        constexpr dequantize_V_t dequantize_k = get_dequantize_V<TYPE_K, half, 4>();
-#pragma unroll
-        for (int i_k = tid; i_k < n_k; i_k += THREADS_PER_BLOCK) {
-            const int i_k_vec = i_k / (N_EMBD / 4);
-            const int i_embd = i_k % (N_EMBD / 4);
-            const int i_kv = start_kv + i_k_vec;
-            if (i_kv < n_kv) {
-                const void * k_base = (const void *) ((const char *) K + i_kv*nbk2 + i_stream*nbk3);
-                dequantize_k(k_base, &k_shared_h[i_k_vec][i_embd][0], i_embd * 4);
-            } else {
-                *(int2*) &k_shared_h[i_k_vec][i_embd] = make_int2(0, 0);
-            }
-        }
-    }
-
-    __syncthreads();
-
-    // phase 3 - calculate lightning indexer scores
-
-    __shared__ float qk_shared[WARPS_PER_BLOCK][HEADS_PER_INNER_LOOP][K_VECS_PER_BLOCK];
-
-    // load K fragment
-    wmma::fragment<wmma::matrix_b, HEADS_PER_INNER_LOOP, K_VECS_PER_BLOCK, K_EMBD_PER_INNER_LOOP, half, wmma::col_major> frag_k;
-    wmma::load_matrix_sync(frag_k, (half*) &k_shared_h[0][i_warp * K_EMBD_PER_INNER_LOOP / 4], N_EMBD_PADDED);
-
-    float score_k = 0.0f;
-
-    for (int i_head_0 = 0; i_head_0 < N_HEAD; i_head_0 += HEADS_PER_INNER_LOOP) {
-        const int i_head_next = i_head_0 + HEADS_PER_INNER_LOOP;
-
-        // we don't use accumulator for anything, fill it with zeros
-        wmma::fragment<wmma::accumulator, HEADS_PER_INNER_LOOP, K_VECS_PER_BLOCK, K_EMBD_PER_INNER_LOOP, float> frag_acc;
-        wmma::fill_fragment(frag_acc, 0.0f);
-
-        // load Q fragment
-        wmma::fragment<wmma::matrix_a, HEADS_PER_INNER_LOOP, K_VECS_PER_BLOCK, K_EMBD_PER_INNER_LOOP, half, wmma::row_major> frag_q;
-        wmma::load_matrix_sync(frag_q, (half*) &q_shared_h[0][i_warp * K_EMBD_PER_INNER_LOOP / 4], N_EMBD_PADDED);
-
-        // preload next Q tile to registers during matrix multiplication
-        float4 q_next[N_Q_NEXT];
-
-        if (i_head_next < N_HEAD) {
-#pragma unroll
-            for (int i_q = tid, i_q_next = 0; i_q < N_Q_TILE; i_q += THREADS_PER_BLOCK) {
-                const int i_head = i_head_next + i_q / (N_EMBD / 4);
-                const int i_embd =               i_q % (N_EMBD / 4);
-                q_next[i_q_next++] = *(const float4 *) (q_base + i_head*nbq1 + i_embd*sizeof(float4));
-            }
-        }
-
-        // perform matrix multiplication
-        wmma::mma_sync(frag_acc, frag_q, frag_k, frag_acc);
-        wmma::store_matrix_sync((float*) &qk_shared[i_warp][0][0], frag_acc, K_VECS_PER_BLOCK, wmma::mem_row_major);
-
-        // make sure all threads finished using q_shared_h so we can store next tile
-        __syncthreads();
-
-        // write preloaded Q tile to shared memory
-        if (i_head_next < N_HEAD) {
-#pragma unroll
-            for (int i_q = tid, i_q_next = 0; i_q < N_Q_TILE; i_q += THREADS_PER_BLOCK) {
-                const int i_head = i_q / (N_EMBD / 4);
-                const int i_embd = i_q % (N_EMBD / 4);
-                half4 q_packed;
-                q_packed.h2[0] = __float22half2_rn(make_float2(q_next[i_q_next].x, q_next[i_q_next].y));
-                q_packed.h2[1] = __float22half2_rn(make_float2(q_next[i_q_next].z, q_next[i_q_next].w));
-                q_shared_h[i_head][i_embd] = q_packed.i2;
-                ++i_q_next;
-            }
-        }
-
-        // accumulate QK multiplication results from all block warps
-        // (there are 256 threads in block and 256 matmul outputs)
-        // TODO it will break if WARP_SIZE is not 32
-        const int h = tid / K_VECS_PER_BLOCK;
-        const int k = tid % K_VECS_PER_BLOCK;
-        const float w_val = w_shared[i_head_0 + h];
-
-        float sum = 0.0f;
-#pragma unroll
-        for (int w = 0; w < WARPS_PER_BLOCK; ++w) {
-            sum += qk_shared[w][h][k];
-        }
-
-        // ReLU, weight
-        sum = sum > 0.0f ? sum : 0.0f;
-        sum *= w_val;
-
-        // wait until qk_shared[0] is no longer used
-        __syncthreads();
-
-        // reuse qk_shared[0] for storing partial results
-        qk_shared[0][h][k] = sum;
-
-        // wait until all threads write their results
-        __syncthreads();
-
-        // accumulate result over heads
-        if (tid < K_VECS_PER_BLOCK) {
-#pragma unroll
-            for (int i_head = 0; i_head < HEADS_PER_INNER_LOOP; ++i_head) {
-                score_k += qk_shared[0][i_head][tid];
-            }
-        }
-
-        // make sure all threads finished using qk_shared
-        __syncthreads();
-    }
-
-    // phase 4 - store output to VRAM
-
-    if (tid < K_VECS_PER_BLOCK) {
-        const int i_kv = start_kv + tid;
-        if (i_kv < n_kv) {
-            const half * m_base = (const half *) ((const char *) M + i_batch*nbm1 + (i_stream%nem3)*nbm3);
-            float * dst_base = (float *) ((char *) dst + i_batch*nb1 + i_stream*nb3);
-            dst_base[i_kv] = score_k + __half2float(m_base[i_kv]);
-        }
-    }
-}
-
-#else // defined(TURING_MMA_AVAILABLE)
-
-template <int WARPS_PER_BLOCK, int K_VECS_PER_BLOCK, int64_t N_EMBD, int64_t N_HEAD, ggml_type TYPE_K>
-static __global__ void lightning_indexer_kernel_wmma(
-        const float * Q, const char * K, const float * W, const half * M, float * dst,
-        int64_t n_stream, int64_t n_batch, int64_t n_kv,
-        size_t nb1, size_t nb2, size_t nb3,
-        size_t nbq1, size_t nbq2, size_t nbq3,
-        size_t nbk1, size_t nbk2, size_t nbk3,
-        size_t nbw1, size_t nbw2, size_t nbw3,
-        size_t nbm1, size_t nbm2, size_t nbm3,
-        int64_t nem3
-    ) {
-    GGML_UNUSED_VARS(Q, K, W, M, dst,
-        n_stream, n_batch, n_kv,
-        nb1, nb2, nb3,
-        nbq1, nbq2, nbq3,
-        nbk1, nbk2, nbk3,
-        nbw1, nbw2, nbw3,
-        nem3);
-    NO_DEVICE_CODE;
-}
-
-#endif // defined(TURING_MMA_AVAILABLE)
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#include "mma.cuh"

 // tokens scored per block by the tile kernel
 #define LIGHTNING_INDEXER_TILE_TOKENS 8
@@ -247,6 +15,173 @@ static __global__ void lightning_indexer_kernel_wmma(
 #define LIGHTNING_INDEXER_TILE_HEADS_PER_PASS 4
 #endif

+
+template <int WARPS_PER_BLOCK, int K_VECS_PER_BLOCK, int64_t N_EMBD, int64_t N_HEAD, ggml_type TYPE_K>
+static __global__ void lightning_indexer_kernel_mma(
+        const float * Q, const char * K, const float * W, const half * M, float * dst,
+        int64_t n_stream, int64_t n_batch, int64_t n_kv,
+        size_t nb1, size_t nb2, size_t nb3,
+        size_t nbq1, size_t nbq2, size_t nbq3,
+        size_t nbk1, size_t nbk2, size_t nbk3,
+        size_t nbw1, size_t nbw2, size_t nbw3,
+        size_t nbm1, size_t nbm2, size_t nbm3,
+        int64_t nem3
+    ) {
+    constexpr int MMA_DIM = 16;
+    constexpr int WAVE_SIZE = ggml_cuda_get_physical_warp_size();
+    constexpr int THREADS_PER_BLOCK = WARPS_PER_BLOCK * WAVE_SIZE;
+    constexpr int N_EMBD_PADDED = N_EMBD + 8;
+    constexpr int K_VECS_PADDED = K_VECS_PER_BLOCK + 8;
+    constexpr int HEAD_CHUNK    = MMA_DIM;
+    constexpr int N_HEAD_CHUNKS = N_HEAD / HEAD_CHUNK;
+    constexpr int N_KV_TILES    = K_VECS_PER_BLOCK / MMA_DIM;
+    constexpr int N_EMBD_TILES  = N_EMBD / MMA_DIM;
+
+    static_assert(N_HEAD           % MMA_DIM == 0, "N_HEAD must be a multiple of 16");
+    static_assert(N_EMBD           % MMA_DIM == 0, "N_EMBD must be a multiple of 16");
+    static_assert(K_VECS_PER_BLOCK % MMA_DIM == 0, "K_VECS_PER_BLOCK must be a multiple of 16");
+
+#if defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
+    using tile_qk = ggml_cuda_mma::tile<MMA_DIM, MMA_DIM, float, ggml_cuda_mma::DATA_LAYOUT_J_MAJOR>;
+#else
+    using tile_qk = ggml_cuda_mma::tile<MMA_DIM, MMA_DIM, float>;
+#endif
+
+    const int i_batch  = blockIdx.y;
+    const int i_stream = blockIdx.z;
+    const int i_warp   = threadIdx.y;
+    const int i_lane   = threadIdx.x;
+    const int tid      = i_warp * WAVE_SIZE + i_lane;
+
+    const int start_kv = blockIdx.x * K_VECS_PER_BLOCK;
+
+    const char  * q_base = (const char  *)                 Q + i_batch*nbq2 + i_stream*nbq3;
+    const float * w_base = (const float *) ((const char *) W + i_batch*nbw1 + i_stream*nbw3);
+
+    __shared__ float w_shared[N_HEAD];
+    __shared__ half  q_shared[2][HEAD_CHUNK][N_EMBD_PADDED];
+    __shared__ half  k_shared[K_VECS_PER_BLOCK][N_EMBD_PADDED];
+    __shared__ float qk_chunk[HEAD_CHUNK][K_VECS_PADDED];
+    __shared__ float score_shared[K_VECS_PER_BLOCK];
+
+    for (int i = tid; i < N_HEAD; i += THREADS_PER_BLOCK) {
+        w_shared[i] = w_base[i];
+    }
+    for (int i = tid; i < K_VECS_PER_BLOCK; i += THREADS_PER_BLOCK) {
+        score_shared[i] = 0.0f;
+    }
+
+    constexpr int N_K_VEC = K_VECS_PER_BLOCK * (N_EMBD / 4);
+    if constexpr (TYPE_K == GGML_TYPE_F16) {
+        for (int i = tid; i < N_K_VEC; i += THREADS_PER_BLOCK) {
+            const int i_k_vec = i / (N_EMBD / 4);
+            const int i_embd4 = i % (N_EMBD / 4);
+            const int i_kv    = start_kv + i_k_vec;
+            if (i_kv < n_kv) {
+                const int2 * k_ptr = (const int2 *) ((const char *) K + i_kv*nbk2 + i_stream*nbk3);
+                *(int2 *) &k_shared[i_k_vec][i_embd4*4] = k_ptr[i_embd4];
+            } else {
+                *(int2 *) &k_shared[i_k_vec][i_embd4*4] = make_int2(0, 0);
+            }
+        }
+    } else {
+        constexpr dequantize_V_t dequantize_k = get_dequantize_V<TYPE_K, half, 4>();
+        for (int i = tid; i < N_K_VEC; i += THREADS_PER_BLOCK) {
+            const int i_k_vec = i / (N_EMBD / 4);
+            const int i_embd4 = i % (N_EMBD / 4);
+            const int i_kv    = start_kv + i_k_vec;
+            if (i_kv < n_kv) {
+                const void * k_ptr = (const void *) ((const char *) K + i_kv*nbk2 + i_stream*nbk3);
+                dequantize_k(k_ptr, &k_shared[i_k_vec][i_embd4*4], i_embd4*4);
+            } else {
+                *(int2 *) &k_shared[i_k_vec][i_embd4*4] = make_int2(0, 0);
+            }
+        }
+    }
+
+    constexpr int N_Q_CHUNK_VEC = HEAD_CHUNK * (N_EMBD / 4);
+    constexpr int N_Q_NEXT = (N_Q_CHUNK_VEC + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
+
+    for (int i = tid; i < N_Q_CHUNK_VEC; i += THREADS_PER_BLOCK) {
+        const int i_head  = i / (N_EMBD / 4);
+        const int i_embd4 = i % (N_EMBD / 4);
+        const float4 q = *(const float4 *) (q_base + i_head*nbq1 + i_embd4*sizeof(float4));
+        const half2 qh[2] = {make_half2(q.x, q.y), make_half2(q.z, q.w)};
+        ggml_cuda_memcpy_1<sizeof(qh)>(&q_shared[0][i_head][i_embd4*4], qh);
+    }
+
+    __syncthreads();
+
+    constexpr ggml_cuda_mma::data_layout input_layout = ggml_cuda_mma::get_input_data_layout();
+
+    for (int hc = 0; hc < N_HEAD_CHUNKS; ++hc) {
+        const int buf  = hc & 1;
+        const int nbuf = (hc + 1) & 1;
+        const bool has_next = (hc + 1) < N_HEAD_CHUNKS;
+
+        float4 q_next[N_Q_NEXT];
+        if (has_next) {
+#pragma unroll
+            for (int i = tid, j = 0; i < N_Q_CHUNK_VEC; i += THREADS_PER_BLOCK, ++j) {
+                const int i_head  = (hc + 1) * HEAD_CHUNK + i / (N_EMBD / 4);
+                const int i_embd4 =                         i % (N_EMBD / 4);
+                q_next[j] = *(const float4 *) (q_base + i_head*nbq1 + i_embd4*sizeof(float4));
+            }
+        }
+
+        for (int kt = i_warp; kt < N_KV_TILES; kt += WARPS_PER_BLOCK) {
+            tile_qk Dqk;
+#pragma unroll
+            for (int et = 0; et < N_EMBD_TILES; ++et) {
+                ggml_cuda_mma::tile<MMA_DIM, MMA_DIM/2, half2, input_layout> Aq;
+                ggml_cuda_mma::tile<MMA_DIM, MMA_DIM/2, half2, input_layout> Bk;
+                ggml_cuda_mma::load_generic(Aq, (const half2 *) &q_shared[buf][0][et*MMA_DIM], N_EMBD_PADDED/2);
+                ggml_cuda_mma::load_generic(Bk, (const half2 *) &k_shared[kt*MMA_DIM][et*MMA_DIM], N_EMBD_PADDED/2);
+                ggml_cuda_mma::mma(Dqk, Aq, Bk);
+            }
+#pragma unroll
+            for (int l = 0; l < Dqk.ne; ++l) {
+                const int head_local = Dqk.get_i(l);
+                const int kv_local   = kt*MMA_DIM + Dqk.get_j(l);
+                qk_chunk[head_local][kv_local] = Dqk.x[l];
+            }
+        }
+
+        __syncthreads();
+
+        if (has_next) {
+#pragma unroll
+            for (int i = tid, j = 0; i < N_Q_CHUNK_VEC; i += THREADS_PER_BLOCK, ++j) {
+                const int i_head  = i / (N_EMBD / 4);
+                const int i_embd4 = i % (N_EMBD / 4);
+                const half2 qh[2] = {make_half2(q_next[j].x, q_next[j].y), make_half2(q_next[j].z, q_next[j].w)};
+                ggml_cuda_memcpy_1<sizeof(qh)>(&q_shared[nbuf][i_head][i_embd4*4], qh);
+            }
+        }
+
+        for (int kv = tid; kv < K_VECS_PER_BLOCK; kv += THREADS_PER_BLOCK) {
+            float s = 0.0f;
+#pragma unroll
+            for (int h = 0; h < HEAD_CHUNK; ++h) {
+                const float qk = qk_chunk[h][kv];
+                s += (qk > 0.0f ? qk : 0.0f) * w_shared[hc*HEAD_CHUNK + h];
+            }
+            score_shared[kv] += s;
+        }
+
+        __syncthreads();
+    }
+
+    for (int i_kv_local = tid; i_kv_local < K_VECS_PER_BLOCK; i_kv_local += THREADS_PER_BLOCK) {
+        const int i_kv = start_kv + i_kv_local;
+        if (i_kv < n_kv) {
+            const half  * m_base   = (const half *) ((const char *) M + i_batch*nbm1 + (i_stream%nem3)*nbm3);
+            float       * dst_base = (float *) ((char *) dst + i_batch*nb1 + i_stream*nb3);
+            dst_base[i_kv] = score_shared[i_kv_local] + __half2float(m_base[i_kv]);
+        }
+    }
+}
+
 // TODO there is one ugly assumption used in this kernel - that WARP_SIZE is equal to 32
 // thanks to that one warp operating on float4 processes whole indexer K/Q vectors
 // 32 * 4 = 128 (N_EMBD)
@@ -616,27 +551,22 @@ void ggml_cuda_lightning_indexer(ggml_backend_cuda_context & ctx, ggml_tensor *
    const int cc     = ggml_cuda_info().devices[device].cc;

    if (n_embd == 128 && n_head == 64) {
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-        if (GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc) && k->type != GGML_TYPE_F32 && k->type != GGML_TYPE_BF16) {
-            // use wmma kernel
+        if ((turing_mma_available(cc) || amd_mfma_available(cc) || amd_wmma_available(cc)) && k->type != GGML_TYPE_F32 && k->type != GGML_TYPE_BF16) {
            constexpr int K_VECS_PER_BLOCK = 32;
-            constexpr int WARPS_PER_BLOCK = 8;
+            constexpr int WARPS_PER_BLOCK  = 4;

-            dim3 block(32, WARPS_PER_BLOCK);
+            dim3 block(ggml_cuda_info().devices[device].warp_size, WARPS_PER_BLOCK);
            int num_kv_blocks = (n_kv + (K_VECS_PER_BLOCK) - 1) / (K_VECS_PER_BLOCK);
            dim3 grid(num_kv_blocks, n_batch, n_stream);

-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 64, k, GGML_TYPE_F16)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 64, k, GGML_TYPE_Q4_0)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 64, k, GGML_TYPE_Q4_1)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 64, k, GGML_TYPE_Q5_0)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 64, k, GGML_TYPE_Q5_1)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 64, k, GGML_TYPE_Q8_0)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 64, k, GGML_TYPE_F16)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 64, k, GGML_TYPE_Q4_0)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 64, k, GGML_TYPE_Q4_1)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 64, k, GGML_TYPE_Q5_0)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 64, k, GGML_TYPE_Q5_1)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 64, k, GGML_TYPE_Q8_0)
            GGML_ABORT("fatal error");
        } else {
-#else // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-        {
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
            // use vector kernel
            constexpr int K_VECS_PER_WARP = 8;
            constexpr int WARPS_PER_BLOCK = 8;
@@ -657,27 +587,22 @@ void ggml_cuda_lightning_indexer(ggml_backend_cuda_context & ctx, ggml_tensor *
            GGML_ABORT("fatal error");
        }
    } else if (n_embd == 128 && n_head == 32) {
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-        if (GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc) && k->type != GGML_TYPE_F32 && k->type != GGML_TYPE_BF16) {
-            // use wmma kernel
+        if ((turing_mma_available(cc) || amd_mfma_available(cc) || amd_wmma_available(cc)) && k->type != GGML_TYPE_F32 && k->type != GGML_TYPE_BF16) {
            constexpr int K_VECS_PER_BLOCK = 32;
-            constexpr int WARPS_PER_BLOCK = 8;
+            constexpr int WARPS_PER_BLOCK  = 4;

-            dim3 block(32, WARPS_PER_BLOCK);
+            dim3 block(ggml_cuda_info().devices[device].warp_size, WARPS_PER_BLOCK);
            int num_kv_blocks = (n_kv + (K_VECS_PER_BLOCK) - 1) / (K_VECS_PER_BLOCK);
            dim3 grid(num_kv_blocks, n_batch, n_stream);

-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 32, k, GGML_TYPE_F16)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 32, k, GGML_TYPE_Q4_0)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 32, k, GGML_TYPE_Q4_1)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 32, k, GGML_TYPE_Q5_0)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 32, k, GGML_TYPE_Q5_1)
-            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_wmma, 128, 32, k, GGML_TYPE_Q8_0)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 32, k, GGML_TYPE_F16)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 32, k, GGML_TYPE_Q4_0)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 32, k, GGML_TYPE_Q4_1)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 32, k, GGML_TYPE_Q5_0)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 32, k, GGML_TYPE_Q5_1)
+            LIGHTNING_INDEXER_CASE(lightning_indexer_kernel_mma, 128, 32, k, GGML_TYPE_Q8_0)
            GGML_ABORT("fatal error");
        } else {
-#else // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-        {
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
            // use vector kernel
            constexpr int K_VECS_PER_WARP = 8;
            constexpr int WARPS_PER_BLOCK = 8;
@@ -76,11 +76,11 @@ namespace ggml_cuda_mma {
        // For the A/C matrices this means I major == row major, J major == column major.
        // For the B matrix this means I major == column major, J major == row major.
        // MIRRORED == Each data value is held exactly once per thread subgroup.
-        DATA_LAYOUT_I_MAJOR           =  0, // Always used for Turing, Ampere, Ada Lovelace, consumer Blackwell, matrix A&B for RDNA4 and CDNA.
-        DATA_LAYOUT_J_MAJOR           = 10, // Matrix C for CDNA and RDNA4, int and float matrix C for RDNA3.
+        DATA_LAYOUT_I_MAJOR           =  0, // Always used for Turing, Ampere, Ada Lovelace, consumer Blackwell, matrix A&B for WMMA128 and CDNA.
+        DATA_LAYOUT_J_MAJOR           = 10, // Matrix C for CDNA and WMMA128, int and float matrix C for WMMA256.
        DATA_LAYOUT_I_MAJOR_MIRRORED  = 20, // Volta, matrix A&B for RDNA3.
        DATA_LAYOUT_J_MAJOR_MIRRORED  = 30,
-        DATA_LAYOUT_I_MAJOR_SCRAMBLED = 40, // Scrambled matrix C for faster transposition (RDNA4/CDNA), convert to float to unscramble.
+        DATA_LAYOUT_I_MAJOR_SCRAMBLED = 40, // Scrambled matrix C for faster transposition (WMMA128/CDNA), convert to float to unscramble.
    };
    // Implemented mma combinations are:
    //   - (I_MAJOR, I_MAJOR)          -> I_MAJOR
@@ -88,11 +88,11 @@ namespace ggml_cuda_mma {
    //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR

    static constexpr __device__ data_layout get_input_data_layout() {
-#if defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
+#if defined(AMD_WMMA_256B_AVAILABLE) || defined(VOLTA_MMA_AVAILABLE)
        return DATA_LAYOUT_I_MAJOR_MIRRORED;
 #else
        return DATA_LAYOUT_I_MAJOR;
-#endif // defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
+#endif // defined(AMD_WMMA_256B_AVAILABLE) || defined(VOLTA_MMA_AVAILABLE)
    }

    template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
@@ -201,7 +201,7 @@ namespace ggml_cuda_mma {

        static __device__ __forceinline__ int get_j(const int l) {
            if constexpr (I == 16 && J == 16) {
-#if defined(RDNA3)
+#if defined(AMD_WMMA_256B_AVAILABLE)
                if constexpr (std::is_same_v<T, float> || std::is_same_v<T, int>) {
                    // matrix C
                    return 2 * l + (threadIdx.x / 16);
@@ -210,11 +210,11 @@ namespace ggml_cuda_mma {
                    return l;
                }
 #else
-                // matrix C is the transposed matrix A&B on RDNA4
+                // matrix C is the transposed matrix A&B on WMMA128
                return ne * (threadIdx.x / 16) + l;
-#endif // defined(RDNA3)
+#endif // defined(AMD_WMMA_256B_AVAILABLE)
            } else if constexpr (I == 16 && J == 8) {
-                // mmq input for RDNA4
+                // mmq input for WMMA128
                return ne * (threadIdx.x / 16) + l;
            } else if constexpr (I == 16 && J == 4) {
                return ne * (threadIdx.x / 16) + l;
@@ -573,7 +573,7 @@ namespace ggml_cuda_mma {
        static constexpr int         I  = I_;
        static constexpr int         J  = J_;
        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR_MIRRORED;
-#if defined(RDNA3)
+#if defined(AMD_WMMA_256B_AVAILABLE)
        static constexpr int         ne = tile<I_, J_, float, DATA_LAYOUT_I_MAJOR_MIRRORED>::ne;

        half2 x[ne] = {{0.0f, 0.0f}};
@@ -616,7 +616,7 @@ namespace ggml_cuda_mma {
                return -1;
            }
        }
-#endif // defined(RDNA3)
+#endif // defined(AMD_WMMA_256B_AVAILABLE)
    };

    template <int I_, int J_>
@@ -694,7 +694,7 @@ namespace ggml_cuda_mma {
    };

    static __device__ __forceinline__ tile<16, 16, half2, DATA_LAYOUT_I_MAJOR> unscramble(const tile<16, 16, half2, DATA_LAYOUT_I_MAJOR_SCRAMBLED> & t) {
-#if defined(AMD_MFMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
        tile<16, 16, half2, DATA_LAYOUT_I_MAJOR> ret;
 #pragma unroll
        for (int l0 = 0; l0 < t.ne/2; ++l0) {
@@ -705,7 +705,7 @@ namespace ggml_cuda_mma {
 #else
        NO_DEVICE_CODE;
        GGML_UNUSED(t);
-#endif // defined(AMD_MFMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
    }

 #if defined(TURING_MMA_AVAILABLE)
@@ -726,7 +726,7 @@ namespace ggml_cuda_mma {

        return ret;
    }
-#elif defined(AMD_WMMA_AVAILABLE) && defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
    static __device__ __forceinline__ tile<16, 8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> get_half2(
            const tile<16, 16, float, DATA_LAYOUT_I_MAJOR> & tile_float) {
        tile<16, 8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> ret;
@@ -1171,7 +1171,7 @@ namespace ggml_cuda_mma {
            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #elif defined(AMD_WMMA_AVAILABLE)
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
        using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
        halfx8_t& acc_frag = reinterpret_cast<halfx8_t&>(D.x[0]);
        const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
@@ -1180,7 +1180,7 @@ namespace ggml_cuda_mma {
 #else
        GGML_UNUSED_VARS(D, A, B);
        NO_DEVICE_CODE;
-#endif // defined(RDNA4)
+#endif // defined(AMD_WMMA_128B_AVAILABLE)
 #elif defined(AMD_MFMA_AVAILABLE)
        // MFMA: FP16 input, FP32 accumulate, convert back to half2.
        using halfx4_t = __attribute__((ext_vector_type(4))) _Float16;
@@ -1218,7 +1218,7 @@ namespace ggml_cuda_mma {
    static __device__ __forceinline__ void mma(
            tile<16, 16, half2, DATA_LAYOUT_I_MAJOR_SCRAMBLED> & D, const tile<32, 8, half2, DATA_LAYOUT_I_MAJOR> & A,
            const tile<16, 8, half2, DATA_LAYOUT_I_MAJOR> & B) {
-#if defined(AMD_MFMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
        tile<16, 8, half2>       * D16 = (tile<16, 8, half2>       *) &D;
        const tile<16, 8, half2> * A16 = (const tile<16, 8, half2> *) &A;
        mma(D16[0], A16[0], B);
@@ -1226,7 +1226,7 @@ namespace ggml_cuda_mma {
 #else
        GGML_UNUSED_VARS(D, A, B);
        NO_DEVICE_CODE;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
    }

    template <data_layout dl_ab, data_layout dl_d>
@@ -1373,14 +1373,14 @@ namespace ggml_cuda_mma {
            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #elif defined(AMD_WMMA_AVAILABLE)
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
        using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
        using floatx8_t = __attribute__((ext_vector_type(8))) float;
        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
        const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
        const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
-#elif defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
        using halfx16_t = __attribute__((ext_vector_type(16))) _Float16;
        using floatx8_t = __attribute__((ext_vector_type(8))) float;
        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
@@ -1390,7 +1390,7 @@ namespace ggml_cuda_mma {
 #else
        GGML_UNUSED_VARS(D, A, B);
        NO_DEVICE_CODE;
-#endif // RDNA4
+#endif // AMD_WMMA_128B_AVAILABLE
 #elif defined(AMD_MFMA_AVAILABLE)
        using halfx4_t = __attribute__((ext_vector_type(4))) _Float16;
        using floatx4_t = __attribute__((ext_vector_type(4))) float;
@@ -1408,14 +1408,14 @@ namespace ggml_cuda_mma {
    static __device__ __forceinline__ void 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)
+#if defined(AMD_WMMA_128B_AVAILABLE)
        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)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
        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]);
@@ -1425,7 +1425,7 @@ namespace ggml_cuda_mma {
 #else
        GGML_UNUSED_VARS(D, A, B);
        NO_DEVICE_CODE;
-#endif // defined(RDNA4)
+#endif // defined(AMD_WMMA_128B_AVAILABLE)
 #elif defined(AMD_MFMA_AVAILABLE)
        using floatx4_t = __attribute__((ext_vector_type(4))) float;
        floatx4_t& acc_frag = reinterpret_cast<floatx4_t&>(D.x[0]);
@@ -1467,19 +1467,19 @@ namespace ggml_cuda_mma {
 #elif defined(AMD_WMMA_AVAILABLE)
        using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
        int32x8_t * acc = (int32x8_t *) D.x;
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
        int32x2_t * a_vec = (int32x2_t *) A.x;
        int32x2_t * b_vec = (int32x2_t *) B.x;
        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], true);
        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[1], true, b_vec[1], acc[0], true);
-#elif defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
        int32x4_t * a_vec = (int32x4_t *) A.x;
        int32x4_t * b_vec = (int32x4_t *) B.x;
        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], true);
        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[1], true, b_vec[1], acc[0], true);
-#endif // RDNA4
+#endif // AMD_WMMA_128B_AVAILABLE
 #else
        GGML_UNUSED_VARS(D, A, B);
        NO_DEVICE_CODE;
@@ -1556,7 +1556,7 @@ namespace ggml_cuda_mma {
    static __device__ __forceinline__ void mma(
            tile<16, 16, half2, DATA_LAYOUT_I_MAJOR> & D, const tile<32,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & A,
            const tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
-#if defined(AMD_WMMA_AVAILABLE) && defined(RDNA3)
+#if defined(AMD_WMMA_256B_AVAILABLE)
        using halfx16_t = __attribute__((ext_vector_type(16))) _Float16;
        halfx16_t       * xD = (halfx16_t       *) D.x;
        const halfx16_t * xA = (const halfx16_t *) A.x;
@@ -1585,17 +1585,17 @@ namespace ggml_cuda_mma {
 #elif defined(AMD_WMMA_AVAILABLE)
        using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
        int32x8_t * acc = (int32x8_t *) D.x;
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
        int32x2_t * a_vec = (int32x2_t *) A.x;
        int32x2_t * b_vec = (int32x2_t *) B.x;
        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], false);
-#elif defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
        int32x4_t * a_vec = (int32x4_t *) A.x;
        int32x4_t * b_vec = (int32x4_t *) B.x;
        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], false);
-#endif // RDNA4
+#endif // AMD_WMMA_128B_AVAILABLE
 #else
        GGML_UNUSED(D);
        GGML_UNUSED(A);
@@ -50,7 +50,8 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
    mmf_ids_data * ids_info_ptr = nullptr;
    ggml_cuda_pool_alloc<int32_t> ids_src_compact_dev;
    ggml_cuda_pool_alloc<int32_t> ids_dst_compact_dev;
-    ggml_cuda_pool_alloc<int32_t> expert_bounds_dev;
+
+    ggml_cuda_pool_alloc<ggml_cuda_mm_id_tile_spec> tile_spec_dev;

    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
    const int64_t ncols_dst          = ids ? ne2  : ne1;
@@ -74,25 +75,28 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
        const int64_t n_experts     = ne02;
        const int64_t n_tokens      = ne12;
        const int64_t ne_get_rows   = n_tokens * n_expert_used;
+        const int64_t max_tiles     = (ne_get_rows + 16 - 1) / 16 + ne02;

        ids_src_compact_dev.alloc(ctx.pool(), ne_get_rows);
        ids_dst_compact_dev.alloc(ctx.pool(), ne_get_rows);
-        expert_bounds_dev.alloc(ctx.pool(), n_experts + 1);
+        tile_spec_dev.alloc(ctx.pool(), max_tiles);

        const int si1  = static_cast<int>(ids_s1);
        const int sis1 = static_cast<int>(src1->nb[2] / src1->nb[1]);

        GGML_ASSERT(sis1 > 0);

-        ggml_cuda_launch_mm_ids_helper(ids_d, ids_src_compact_dev.get(), ids_dst_compact_dev.get(), expert_bounds_dev.get(),
-            static_cast<int>(n_experts), static_cast<int>(n_tokens), static_cast<int>(n_expert_used), static_cast<int>(ne11), si1, sis1, /*write_inverse =*/ false, ctx.stream());
+        ggml_cuda_launch_mm_ids_helper(ids_d, ids_src_compact_dev.get(), ids_dst_compact_dev.get(), tile_spec_dev.get(),
+            static_cast<int>(n_experts), static_cast<int>(n_tokens), static_cast<int>(n_expert_used), static_cast<int>(ne11), 16,
+            si1, sis1, /*write_inverse =*/ false, ctx.stream());
        CUDA_CHECK(cudaGetLastError());

-        ids_info.ids_src_compact   = ids_src_compact_dev.get();
-        ids_info.ids_dst_compact   = ids_dst_compact_dev.get();
-        ids_info.expert_bounds_dev = expert_bounds_dev.get();
-        ids_info.n_experts         = static_cast<int>(n_experts);
-        ids_info.sis1              = sis1;
+        ids_info.ids_src_compact = ids_src_compact_dev.get();
+        ids_info.ids_dst_compact = ids_dst_compact_dev.get();
+        ids_info.tile_spec_dev   = tile_spec_dev.get();
+        ids_info.n_experts       = static_cast<int>(n_experts);
+        ids_info.max_tiles       = static_cast<int>(max_tiles);
+        ids_info.sis1            = sis1;
        ids_info_ptr = &ids_info;
    }

@@ -1,7 +1,8 @@
 #pragma once

-#include "mma.cuh"
 #include "common.cuh"
+#include "mma.cuh"
+#include "mmid.cuh"
 #include "convert.cuh"

 using namespace ggml_cuda_mma;
@@ -36,8 +37,9 @@ static constexpr __device__ int mmf_get_padding() {
 struct mmf_ids_data {
    const int32_t * ids_src_compact = nullptr;
    const int32_t * ids_dst_compact = nullptr;
-    const int32_t * expert_bounds_dev = nullptr;
+    const ggml_cuda_mm_id_tile_spec * tile_spec_dev = nullptr;
    int n_experts = 0;
+    int max_tiles = 0;
    int sis1 = 0;
 };

@@ -308,7 +310,7 @@ __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
 static __global__ void mul_mat_f_ids(
        const T * __restrict__ x, const float * __restrict__ y,
        const int32_t * __restrict__ ids_src_compact, const int32_t * __restrict__ ids_dst_compact,
-        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst,
+        const ggml_cuda_mm_id_tile_spec * __restrict__ tile_spec, float * __restrict__ dst,
        const int ncols, const int ncols_dst_total, const int nchannels_dst, const int stride_row, const int stride_col_y, const int stride_col_dst,
        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
@@ -349,21 +351,20 @@ static __global__ void mul_mat_f_ids(
    constexpr int ntA = rows_per_block / tile_A::I;
    constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;

-    const int row0        = blockIdx.x * rows_per_block;
+    const int row0     = blockIdx.x * rows_per_block;
+    const int tile_idx = blockIdx.y;

-    const int expert_idx = blockIdx.y;
-    const int expert_start = expert_bounds[expert_idx];
-    const int expert_end   = expert_bounds[expert_idx + 1];
+    const ggml_cuda_mm_id_tile_spec ts = tile_spec[tile_idx];
+
+    const int expert_idx   = ts.expert;
+    const int expert_start = ts.col_low;
+    const int expert_end   = ts.col_high;
    const int ncols_expert = expert_end - expert_start;

-    const int tiles_for_expert = (ncols_expert + cols_per_block - 1) / cols_per_block;
-    const int tile_idx = blockIdx.z;
-    if (tile_idx >= tiles_for_expert) {
+    if (ncols_expert == 0) {
        return;
    }

-    const int col_base = tile_idx * cols_per_block;
-
    GGML_UNUSED(channel_ratio);

    const int channel_x   = expert_idx;
@@ -409,10 +410,9 @@ static __global__ void mul_mat_f_ids(
 #pragma unroll
                for (int j0 = 0; j0 < tile_B::I; ++j0) {
                    const int j = j0 + tile_idx_local*tile_B::I;
-                    const int global_j = col_base + j;
                    float val = 0.0f;
-                    if (j < cols_per_block && global_j < ncols_expert) {
-                        const int src_entry = ids_src_expert[global_j];
+                    if (j < cols_per_block && j < ncols_expert) {
+                        const int src_entry = ids_src_expert[j];
                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
                        const int token   = (int) qrm.x;
                        const int channel = (int) qrm.y;
@@ -462,10 +462,9 @@ static __global__ void mul_mat_f_ids(
 #pragma unroll
                for (int j0 = 0; j0 < tile_B::I; ++j0) {
                    const int j = j0 + tile_idx_local*tile_B::I;
-                    const int global_j = col_base + j;
                    float2 tmp = make_float2(0.0f, 0.0f);
-                    if (j < cols_per_block && global_j < ncols_expert) {
-                        const int src_entry = ids_src_expert[global_j];
+                    if (j < cols_per_block && j < ncols_expert) {
+                        const int src_entry = ids_src_expert[j];
                        const uint2 qrm = fast_div_modulo((uint32_t) src_entry, sis1_fd);
                        const int token   = (int) qrm.x;
                        const int channel = (int) qrm.y;
@@ -564,9 +563,8 @@ static __global__ void mul_mat_f_ids(
            }
        }

-        const int global_j = col_base + j;
-        if (j < cols_per_block && global_j < ncols_expert && nchannels_dst > 0) {
-            const int dst_entry = ids_dst_expert[global_j];
+        if (j < cols_per_block && j < ncols_expert && nchannels_dst > 0) {
+            const int dst_entry = ids_dst_expert[j];
            const uint2 qrm = fast_div_modulo((uint32_t) dst_entry, nch_fd);
            const int token = (int) qrm.x;
            if (token < ncols_dst_total) {
@@ -580,7 +578,7 @@ static __global__ void mul_mat_f_ids(
    }
    }
 #else
-    GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst,
+    GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, tile_spec, dst,
        ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, sis1_fd, nch_fd);
@@ -603,17 +601,16 @@ static inline void mul_mat_f_switch_ids(
    // Use the compact-ids kernel only for larger tiles; for small ncols_dst (< 16)
    // we prefer the normal mul_mat_f path with has_ids=true.
    if (has_ids_data && ncols_dst > 16) {
-        const int max_tiles = (int) ((ncols_dst + cols_per_block - 1) / cols_per_block);
-        if (max_tiles == 0) {
+        if (ids_data->max_tiles == 0) {
            return;
        }
-        dim3 block_nums_ids(block_nums.x, ids_data->n_experts, max_tiles);
+        dim3 block_nums_ids(block_nums.x, ids_data->max_tiles, 1);

        const uint3 sis1_fd = ids_data->sis1 > 0 ? init_fastdiv_values((uint32_t) ids_data->sis1) : make_uint3(0, 0, 1);
        const uint3 nch_fd  = init_fastdiv_values((uint32_t) nchannels_dst);

        mul_mat_f_ids<T, rows_per_block, cols_per_block, nwarps><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
-            (x, y, ids_data->ids_src_compact, ids_data->ids_dst_compact, ids_data->expert_bounds_dev, dst,
+            (x, y, ids_data->ids_src_compact, ids_data->ids_dst_compact, ids_data->tile_spec_dev, dst,
            ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
            channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
            sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst,
@@ -27,12 +27,16 @@ template <>      struct mm_ids_pow2<0> { static constexpr int value = 1; };
 // Helper function for mul_mat_id, converts ids to a more convenient format.
 // ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
 // ids_dst describes the same mapping but for the dst tensor.
-// The upper and lower bounds for the ith expert in the compact src1 tensor are stored in expert_bounds[i:i+1].
+// The first (n_tokens*n_expert_used + J - 1) / tile_width elements of tile_spec contain the specication for
+//   matrix multiplication tiles at token positions which are exactly divided by tile_width.
+// Additionally the following n_expert elements of tile_spec may contain additional tiles that fall between these seams.
 template <int n_expert_used_template>
 __launch_bounds__(ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mm_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, const bool write_inverse) {
+        const int32_t * __restrict__ ids,
+        int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, ggml_cuda_mm_id_tile_spec * __restrict__ tile_spec,
+        const int n_tokens, const int n_expert_used_var, const int nchannels_y, const uint3 tile_width,
+        const int si1, const int sis1, const bool write_inverse) {
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
    const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
    const int expert = blockIdx.x;
@@ -118,19 +122,29 @@ static __global__ void mm_ids_helper(
        return;
    }

-    expert_bounds[expert] = nex_prev;
+    int col_low  = nex_prev;
+    int col_high = nex_prev + min(int(tile_width.z), it_compact);

-    if (expert < static_cast<int>(gridDim.x) - 1) {
-        return;
+    // The while loop covers all tiles at the seams containing the column with % tile_width == 0.
+    while (col_high > fastdiv(col_low + tile_width.z - 1, tile_width) * tile_width.z) {
+        tile_spec[fastdiv(col_low + tile_width.z - 1, tile_width)]
+            = {expert, col_low, col_high, /*padding =*/ 0};
+
+        col_low  = col_high;
+        col_high = min(col_high + int(tile_width.z), nex_prev + it_compact);
    }

-    expert_bounds[gridDim.x] = nex_prev + it_compact;
+    // Finally, at most one extra tile_spec between the seams per expert.
+    tile_spec[fastdiv(n_tokens*n_expert_used + tile_width.z - 1, tile_width) + expert]
+        = {expert, col_low, col_high, /*padding =*/ 0};
 }

 template <int n_expert_used_template>
 static void launch_mm_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int si1, const int sis1, const bool write_inverse, cudaStream_t stream) {
+        const int32_t * __restrict__ ids,
+        int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, ggml_cuda_mm_id_tile_spec * __restrict__ tile_spec,
+        const int n_experts, const int n_tokens, const int n_expert_used_var, const int nchannels_y, const int tile_width,
+        const int si1, const int sis1, const bool write_inverse, cudaStream_t stream) {
    GGML_ASSERT(n_tokens          < (1 << 22) && "too few bits in mm_ids_helper_store");
    GGML_ASSERT(n_expert_used_var < (1 << 10) && "too few bits in mm_ids_helper_store");

@@ -139,41 +153,53 @@ static void launch_mm_ids_helper(
    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
    CUDA_SET_SHARED_MEMORY_LIMIT(mm_ids_helper<n_expert_used_template>, smpbo);

+    const uint3 tile_width_fd = init_fastdiv_values(tile_width);
+
    const dim3 num_blocks(n_experts, 1, 1);
    const dim3 block_size(warp_size, 1, 1);
    const size_t nbytes_shared = n_tokens*sizeof(mm_ids_helper_store);
    GGML_ASSERT(nbytes_shared <= smpbo);
    mm_ids_helper<n_expert_used_template><<<num_blocks, block_size, nbytes_shared, stream>>>
-        (ids, ids_src1, ids_dst, expert_bounds, n_tokens, n_expert_used_var, nchannels_y, si1, sis1, write_inverse);
+        (ids, ids_src1, ids_dst, tile_spec, n_tokens, n_expert_used_var, nchannels_y, tile_width_fd, si1, sis1, write_inverse);
 }

 void ggml_cuda_launch_mm_ids_helper(
-        const int32_t * __restrict__ ids, int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, int32_t * __restrict__ expert_bounds,
-        const int n_experts, const int n_tokens, const int n_expert_used, const int nchannels_y, const int si1, const int sis1, const bool write_inverse, cudaStream_t stream) {
+        const int32_t * __restrict__ ids,
+        int32_t * __restrict__ ids_src1, int32_t * __restrict__ ids_dst, ggml_cuda_mm_id_tile_spec * __restrict__ tile_spec,
+        const int n_experts, const int n_tokens, const int n_expert_used, const int nchannels_y, const int tile_width,
+        const int si1, const int sis1, const bool write_inverse, cudaStream_t stream) {
    switch (n_expert_used) {
        case  2:
-            launch_mm_ids_helper< 2>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper< 2>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
        case  4:
-            launch_mm_ids_helper< 4>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper< 4>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
        case  6:
-            launch_mm_ids_helper< 6>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper< 6>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
        case  8:
-            launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
        case 10:
-            launch_mm_ids_helper<10>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper<10>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
        case 16:
-            launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
        case 32:
-            launch_mm_ids_helper<32>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper<32>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
        default:
-            launch_mm_ids_helper< 0>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
+            launch_mm_ids_helper< 0>(ids, ids_src1, ids_dst, tile_spec,
+                n_experts, n_tokens, n_expert_used, nchannels_y, tile_width, si1, sis1, write_inverse, stream);
            break;
    }
 }
--- a/Show More
+++ b/Show More