Image-to-Image
Diffusers
English
art

Fail to load lora weight with pipe.load_lora_weights("./ICEdit-MoE-LoRA/ICEdit-MoE-LoRA.safetensors")

#2
by idonashino - opened

Dear sanaka:

Thank you for your awesome work! But when I try to use it with python

pipe = FluxFillPipeline.from_pretrained("./FLUX.1-Fill-dev", torch_dtype=torch.bfloat16)
pipe.load_lora_weights("./ICEdit-MoE-LoRA/ICEdit-MoE-LoRA.safetensors")
pipe.enable_model_cpu_offload() 

the terminal outputs the following errors:

NotImplementedError: Cannot copy out of meta tensor; no data! Please use torch.nn.Module.to_empty() instead of torch.nn.Module.to() when moving module from meta to a different device.

But when I use this:

pipe.load_lora_weights("RiverZ/normal-lora")

it works.

It seems like I didn't load the lora weight in a correct way. Could you please help me? Many thanks for your kindness!

I'm very sorry for the inconvenience caused. Our MoE LoRA needs to be used with the modified versions of the PEFT library and Diffusers library from our repository. However, due to some unfortunate reasons, we were forced to temporarily remove the scripts for these two libraries. You can refer to repositories that others have forked before, such as https://github.com/judian17/ICEdit, which includes our scripts.

Once again, I sincerely apologize and thank you for your understanding!

I'm very sorry for the inconvenience caused. Our MoE LoRA needs to be used with the modified versions of the PEFT library and Diffusers library from our repository. However, due to some unfortunate reasons, we were forced to temporarily remove the scripts for these two libraries. You can refer to repositories that others have forked before, such as https://github.com/judian17/ICEdit, which includes our scripts.

Once again, I sincerely apologize and thank you for your understanding!

Great! It works for me now! Thank you :D

Sign up or log in to comment