Bug : there is a bug in the finetune.py, that would result to an empyt adapter checkpoint after training!!!

[DragonMengLong]

In the finetune.py, lin line 246, where the original model.state_dict is changed into a lambda function of get_peft_model_state_dict()

get_peft_model_state_dict() will retuen a dict of only the Lora parameter and the keys of the parameter is regularized, the adapter_name in the key string is removed.
to_return = {k.replace(f".{adapter_name}", ""): v for k, v in to_return.items()}

Then when saving the model after Training, In the finetune.py line 275, use the model.save_pretrained() function, which is implemented by the PeftModel.save_pretrained(). Inside the PeftModel.save_pretrained(), get_peft_model_state_dict() is used again.

However, because the model.state_dict is already changed! The dict returened by it is regularized (changed keys' name), the the get_peft_model_state_dict() inside the PeftModel.save_pretrained() will remove all the parameter whose keys string doesn't contain the adapter_name, which is all! It will result in the final state_dict to be save is None!!!
to_return = {k: v for k, v in to_return.items() if (("lora_" in k and adapter_name in k) or ("bias" in k))}

[TianqiYe]

yea, I just realized that this morning after all-night training lol

[minruigui]

I've identified the issue. It lies within the code:

old_state_dict = model.state_dict
model.state_dict = (
lambda self, *_, **__: get_peft_model_state_dict(
self, old_state_dict()
)
).get(model, type(model))

Commenting out this part will resolve the issue."

[vihangd]

I have tried to fix some of these issues in my QLoRA fork of alpaca-lora https://github.com/vihangd/alpaca-qlora Feel free to give it a go and report any issues that you might encounter.

[cwzhao]

I just delete the adapter_name in k in perf. No idea if it will cause some other bugs