Some questions about enabling deepspeed in cogvideo training scripts

[luyvlei]

In this 1448 line of this script, when deepspeed is enabled, optimizer.step() and zero_grad() will no longer be executed. I have checked the related documentation of accelerate and found no similar introduction, is this correct?

diffusers/examples/cogvideo/train_cogvideox_lora.py

Lines 1448 to 1450 in d9029f2

	if accelerator.state.deepspeed_plugin is None:
	optimizer.step()
	optimizer.zero_grad()

[a-r-r-o-w]

The optimizer step and gradient zero-ing is handled internally within the deepspeed engine I think (not 100% sure), so it is okay to skip. I believe even if executed, these will be no-ops. I think to reduce confusion, we could do these but I don't have the bandwidth to test on a complete run. Would you be able to verify? cc @yiyixuxu