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

Thanks! It solves the bug of SD pipeline, but seems like SDXL pipeline is still not working.
Colab notebook

Nice, that's an important fix! Let's merge 🚀

BTW this bug might be solved natively on Torch 2.1 now.

Also cc @younesbelkada @pacman100 here. Having torch.compile work with PEFT + diffusers would be important

@ziniuwan weird that it happens in specific cases. For example, if you call the pipeline (where the UNet is already compiled) and then load the LoRA params and call the pipeline, it works:

https://colab.research.google.com/gist/sayakpaul/58d9640998be5e1852ca630a7b006a8c/reproduce_lora_sdxl_bug.ipynb

@patrickvonplaten have you faced #5298 (comment) before?

@ziniuwan we're merging this PR. For SDXL, we suggest using fuse_lora() to achieve speedup (this is also a general suggestion in case you want to speed up the LoRA computations).

fuse_lora() docs:

https://huggingface.co/docs/diffusers/main/en/training/lora#fusing-lora-parameters