pipe.to("cuda") runs super slow, is it expected?
The pipe = FluxKontextPipeline.from_pretrained(..., torch_dtype=torch.bfloat16) takes a few minutes, while pipe.to("cuda") takes about half an hour (or even more). I'm running the code on an nvidia-h100-80g card.
Is this normal? Do we have some way to accelerate this procedure?
update: it seems only the first time of to("cuda") takes a lot of time. Run the code again and it only takes seconds. Though I'm not sure what the cause is
On my H100 card it doesn't take too much time. Could you share what diffusers version you're using? Also, since this issue seems to be related to the diffusers library, could you please close this discussion and open an issue to the Diffusers repository, instead?
On my H100 card it doesn't take too much time. Could you share what
diffusersversion you're using? Also, since this issue seems to be related to thediffuserslibrary, could you please close this discussion and open an issue to the Diffusers repository, instead?
the version is diffusers-0.35.0.dev0