SageAttention

TSAIL's plug-and-play quantized attention (INT8 → FP4 across v1/2/3; 3.5K★), 2–5× faster than FlashAttention and the de-facto standard in open video-generation inference (ComfyUI, HunyuanVideo, Wan, CogVideoX). The line continues with SpargeAttention (ICML 2025) and TurboDiffusion's 100–200× video-diffusion acceleration (3.6K★ in 7 months).