ShreejithSG
dfaa35de41
- Unify sliding_tile_attn, video_sparse_attn, and vmoba into a standalone, installable package - Simplify build process with a single setup.py for CUDA extensions and Triton fallbacks - Optimize for H100 architectures using ThunderKittens - Validated implementation correctness against PyTorch flex_attention on H100 hardware
FastVideo Kernel
CUDA kernels for FastVideo video generation.
Installation
git submodule update --init --recursive
cd csrc/fastvideo_kernel
pip install .
Usage
from fastvideo_kernel import sliding_tile_attention, video_sparse_attn
Requirements
- H100 GPU (sm_90a) for CUDA kernels
- Triton for non-H100 fallback