arxivcs.CV2026-07-17
FVAttn: Adaptive Sparse Attention with Runtime Load Balancing for Video Generation
Hao Liu, Chenghuan Huang, Ye Huang, Zhiying Wen, Hao Liu, Mohan Zhang, et al.
Video Diffusion Transformers process long spatio-temporal sequences, making self-attention the main bottleneck in high-resolution video generation. Training-free sparse attention reduces this cost, but adaptive Top-$p$ routing creates uneven per-head workloads under multi-GPU seq…