Is there any deepspeed support?

[KleinXin]

It is so memory consuming especially during the starting phase.

Is it possible to use deepspeed?

thx

[TianxiangMa]

Yes, this model is trained and inferred based on the original Wan2.1 model, and the inference code matches the official Wan2.1 implementation. As a result, it does consume a significant amount of GPU memory. However, if there are any acceleration or memory optimization strategies developed for Wan2.1 inference, they can also be applied to our model.