use xformers on linux with diffusers==0.13.0

[chengziyi]

when i run this code on linux with xformers
it reports ERROR: "forward_xformers() got an unexpected keyword argument 'encoder_hidden_states'"

i found a way to avoid it, not the elegant way, but can work with diffuser==0.13.0
python==3.9 , torch==1.13.1, xformers==0.0.16rc425

1. find function: replace_unet_cross_attn_to_xformers() in library/train_util.py
   remove this line: "diffusers.models.attention.CrossAttention.forward = forward_xformers"

2.open site-packages/diffusers/models/attention.py, find "class BasicTransformerBlock", add "self.attn1.set_use_memory_efficient_attention_xformers(True)" after self.attn1 = CrossAttention(***)

[CHR-ray]

Thank you, I have reinstall several version of xformers, torch, cuda...
None of them really work, only yours.
My setting:
python 3.10
torch 1.13.1
torchvision 0.14.1
diffusers 0.13.1
xformers 0.0.16

[CHR-ray]

sorry, I reply too early. When I restart my lora script, it even raise out of vram at 1024 * 1024.... (24GB VRAM I have)

Maybe there are still something wrong...

[chengziyi]

PS: what i do is avoid this ERROR: "forward_xformers() got an unexpected keyword argument 'encoder_hidden_states'" with diffusers==0.13 on linux. Obviously it is not a solution to solve the problem