Sage Attention vs Flash Attention Speed Comparison with Wan 2.1 - Sage Attention is 37% Faster - 720p - 14b model - tested on Windows Python VENV - no WSL

I am developing a Wan 2.1 Gradio App

Using DiffSynth-Studio as a backend - I am using slightly modified back end though you can see my pull request here : https://github.com/modelscope/DiffSynth-Studio/pull/463

I just implemented Sage Attention to my app since Windows pre-compiled wheels published by a hero

Here comparison. I feel like just a little bit quality degrade but huge speed up. Tested on RTX 5090 on Windows with native Python 3.10 VENV - (no WSL)

https://github.com/user-attachments/assets/59df3dc1-b5cc-498d-90fc-d83782ec75bc

Here my app link : https://www.patreon.com/posts/123105403

The app is 1-click to install with all pre-compiled wheels

Here my app interface

Prompt

Close-up shot of a smiling young boy with a joyful expression, sitting comfortably in a cozy room. The boy has tousled brown hair and wears a colorful t-shirt. Bright, soft lighting highlights his happy face. Medium close-up, slightly tilted camera angle.

Negative Prompt

Overexposure, static, blurred details, subtitles, paintings, pictures, still, overall gray, worst quality, low quality, JPEG compression residue, ugly, mutilated, redundant fingers, poorly painted hands, poorly painted faces, deformed, disfigured, deformed limbs, fused fingers, cluttered background, three legs, a lot of people in the background, upside down

The issue contains a performance comparison and links to a related DiffSynth-Studio pull request; start by reviewing that pull request and the Sage Attention integration it describes. No repository files or tests are named, and there is no requested change or acceptance criterion, so a concrete goal and validation method would need to be established before work can be considered done.