Exception: Error while deserializing header: HeaderTooLarge

[jelling]

Describe the bug

There may be multiple reasons for the Exception: Error while deserializing header: HeaderTooLarge error message when loading a LoRA, but one of them is caused by passing in /pytorch_lora_weights.bin instead of just the root directory.

Suggestion: add a check to make certain the user only passes in a directory OR drop the filename part. Also, the LoRA docs currently do not show a full path example which is probably why I hit this.

Reproduction

Example:

  # Add LoRA
  bad_path = "/some_lora_dir/pytorch_lora_weights.bin"
  pipe.load_lora_weights(lora_path) #  Error while deserializing header: HeaderTooLarge
  lora_path = = "/some_lora_dir" 
  pipe.load_lora_weights(good_path) # works fine

Logs

No response

System Info

Doesn't matter

[patrickvonplaten]

Agree a better error message would be nice here! @sayakpaul do you have time to look into this maybe?

[sayakpaul]

Is it not clear from the docstring?

https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights

Also, the LoRA docs currently do not show a full path example which is probably why I hit this.

What is a full path?

[jelling]

Is it not clear from the docstring?

https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights

Yes, the API docs are clear. I should have said the LoRA example page. And by full path, I meant a full local file path.

But since most people will likely be using HF as their source for LoRAs, perhaps just adding a "See also:" section to the example with links to the API docs would be helpful. But this is a nit-pick, the docs are overall solid.

[sayakpaul]

Thanks @jelling.

I opened a PR to clarify that: #3397.

[Mineyme]

I also have the similar issue when I run my comfyUI On the cloud。I cannot find any file about lora ,except /notebooks/ComfyUI/models/loras. can anyone tell me how to resolve it, I'm not a developer , detail as soon as possible ,thank you very much