README.md

LightX2V NVIDIA Kernels

The wheel distribution is lightx2v-kernel-nvidia, imported as lightx2v_kernel_nvidia. LightX2V uses the new package when installed and falls back to the legacy lightx2v_kernel package when it is absent, so existing wheels can still be used without rebuilding. The legacy wheel must include the operators required by the selected model.

The C++ operator namespace remains torch.ops.lightx2v_kernel for binary compatibility.

Preparation

# Install torch, at least version 2.7

pip install scikit_build_core uv

Build whl

The default target is 120a; for H100, add -Ccmake.define.LIGHTX2V_GPU_ARCH=90a to the build command.

The Hopper build also provides FP8 GEMM with FP32 fast accumulation. H3 can select it with dit_quant_scheme / video_vae_quant_scheme: "fp8-f32-accum", reusing the fp8-f16-accum checkpoints, quantization ranges, and fused epilogue. The original fp8-f16-accum mode retains FP16 accumulation.

git clone https://github.com/NVIDIA/cutlass.git

git clone https://github.com/ModelTC/LightX2V.git

cd LightX2V/lightx2v_kernel/nvidia

# Set the /path/to/cutlass below to the absolute path of cutlass you download.

MAX_JOBS=$(nproc) && CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) \
uv build --wheel \
    -Cbuild-dir=build . \
    -Ccmake.define.CUTLASS_PATH=/path/to/cutlass \
    --verbose \
    --color=always \
    --no-build-isolation

Install whl

pip install dist/*whl --force-reinstall --no-deps

Test

cos and speed test, mm without bias

python test/nvfp4_nvfp4/test_bench2.py

cos and speed test, mm with bias

python test/nvfp4_nvfp4/test_bench3_bias.py

Bandwidth utilization test for quant

python test/nvfp4_nvfp4/test_quant_mem_utils.py

tflops test for mm

python test/nvfp4_nvfp4/test_mm_tflops.py