Using --use-sage-attention did not result in any speed improvement

[alboto]

docker image: nvidia/cuda:12.6.3-cudnn-devel-ubuntu22.04
pytorch version: 2.7.0.dev20250105+cu126
python version: 3.10.0
nvidia device: L20
Driver Version: 560.35.03
comfyui commit: a90aafa
workflow: test (1).json

Could it be because I made a mistake somewhere?

[blepping]

You probably won't notice of a difference with a 768x576 generation, these performance optimizations typically make the most difference at fairly high resolutions. It also depends on what the bottleneck is for generation, if most of the time is spent shuffling stuff back and forth (VRAM inadequate) then making the computation faster probably won't help that much.

[Colinw2292]

Might be a silly question but did you install Sage Attention? If so what version?

[blepping]

ComfyUI doesn't check if you actually have SageAttention, so if you do --use-sage-attention without it you'll get a crash as soon as any model tries to use attention.