Different running time between Cosmos Framework and Diffusers

[TriCaoChanh]

Hi, I notice a significant increasing of running time when perform inference with Diffusers on Cosmos-Nano, following the inference guide.

Is is the expected behavior? What could be the problem and any idea for speeding up?

Cosmos-Nano running time:

• Diffusers: 5 minutes 46s
• Cosmos Framework: 3 minutes 48s

[foreverlms]

Thanks for this finding, we will investigate on it soon!

[foreverlms]

We identified that by default, cosmos framework use attention backend from tranformer engine; but diffusers uses native pytorch SDPA. That may be the cause.