[Dtype] Align dtype casting behavior with Transformers and Accelerate

Open patrickvonplaten opened this issue 3 years ago • 4 comments

Some recent changes in transformers: https://github.com/huggingface/transformers/pull/20602 and accelerate: https://github.com/huggingface/accelerate/pull/920 that force us to also align the behavior in diffusers. For more information also have a look at: https://discuss.pytorch.org/t/discrepancy-between-loading-models-with-meta-tensors-and-normal-load-from-state-dict/168295

Dec 16 '22 11:12 patrickvonplaten

The documentation is not available anymore as the PR was closed or merged.

Dec 16 '22 11:12 HuggingFaceDocBuilderDev

This pretty much reverses: https://github.com/huggingface/diffusers/pull/1449

Dec 16 '22 11:12 patrickvonplaten

Wait until https://github.com/huggingface/accelerate/pull/920 is merged.

Dec 16 '22 12:12 patrickvonplaten

@pcuenca @patil-suraj feel free to merge whenever. Maybe a nice message here explaining what changed would make sense as well

Dec 20 '22 17:12 patrickvonplaten