AB-Attention Speeds Up Video Diffusion Transformers with Training-Free Accurate Block-Sparse Attention
Bin Xu · hf · 2026-10-30
A multi-institution team proposes AB-Attention, a training-free method that approximates dense 3D attention in video diffusion transformers at high sparsity. Its Intra-Block Clustering Selector estimates block importance cheaply by clustering queries only within each query block without permuting tokens, addressing fidelity loss of prior block-sparse methods under aggressive sparsity, complemented by weighted sampling and fine-grained kernel gathering for both speed and fidelity gains.
More from Infra
- RTX 5080 Pulled from NVIDIA Marketplace; 24GB Variant Rumored for Early 2027 — chemist_slime · 2026-10-30
- Amazon invests another $5B in Anthropic, securing $100B+ decade-long AWS commitment — emmanuelvivier · 2026-10-30
- Nvidia unveils Rubin: six new chips and a rack-scale AI supercomputer — emmanuelvivier · 2026-10-30
- Amazon and Intel earnings misses deepen AI bubble fears, tech stocks slide — nordicinst · 2026-10-30
- NVIDIA RTX Spark pairs Grace CPU with Blackwell GPU sharing up to 128GB, 1 petaflop FP4 — bestvicam · 2026-10-30
- Ollama installed ≠ model downloaded: pitfalls when running it on cloud GPUs — NoConflict2854 · 2026-10-30