mit-han-lab/radial-attention

Implements physics-inspired sparse attention masks with exponentially decaying compute density across temporal bands, integrating with video diffusion models (Wan2.1, HunyuanVideo, Mochi-1) and optimized backends including SageAttention and FlashInfer. Achieves O(n log n) complexity through static spatiotemporal masking that scales pre-trained models to 4× longer sequences via lightweight LoRA tuning, with multi-GPU support via xDiT's Ulysses sequence parallelism.

587 stars.