Bug when pass subfolder parameter in peftmodel.load_adapter method!

[ShayDuane]

System Info

peft 0.4.0.dev0

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder
• My own task or dataset (give details below)

Reproduction

Given a checkpoint folder as shown below.

drwxr-xr-x 2 root root     4096 Jul 16 09:54 LoraAdapter
-rw-r--r-- 1 root root 23443269 Jul 16 09:54 optimizer.pt
-rw-r--r-- 1 root root       27 Jul 16 09:54 README.md
-rw-r--r-- 1 root root    14575 Jul 16 09:54 rng_state.pth
-rw-r--r-- 1 root root      627 Jul 16 09:54 scheduler.pt
-rw-r--r-- 1 root root     7592 Jul 16 09:54 trainer_state.json
-rw-r--r-- 1 root root     3963 Jul 16 09:54 training_args.bin

LoraAdapter is the name I specified for the Lora adapter, which is a folder here containing the weight files and configuration files of the adapter.

-rw-r--r-- 1 root root      411 Jul 16 09:54 adapter_config.json
-rw-r--r-- 1 root root 11719577 Jul 16 09:54 adapter_model.bin

I want to continue training from this checkpoint, so I need to instantiate a PeftModel and load the weights of the Lora adapter into it. However, my adapter weight file is in a subfolder of the checkpoint, so I use the following method.

from transformers import AutoModel
from peft import PeftModelForCausalLM, LoraConfig

lora_config = LoraConfig.from_pretrained('ChatGlm2-6B_LoRA/QLora_Adapter_THUDM_chatglm2-6b/checkpoint-2000', subfolder='LoraAdapter')
base_model = AutoModel.from_pretrained('THUDM/chatglm2-6b', trust_remote_code=True, device_map='auto').half()
peftmodel = PeftModelForCausalLM(base_model, lora_config, adapter_name='LoraAdapter')

peftmodel.load_adapter(model_id='ChatGlm2-6B_LoRA/QLora_Adapter_THUDM_chatglm2-6b/checkpoint-2000', adapter_name='LoraAdapter', subfolder='LoraAdapter')

But it cannot correctly find my Lora adapter weight file.

I think the problem is here.

peft/src/peft/peft_model.py

Lines 484 to 502 in 029f416

	def _split_kwargs(cls, kwargs: Dict[str, Any]):
	_kwargs_not_in_hf_hub_download_signature = ("use_auth_token",)
	hf_hub_download_kwargs = {}
	other_kwargs = {}
	for key, value in kwargs.items():
	if key in inspect.signature(hf_hub_download).parameters or key in _kwargs_not_in_hf_hub_download_signature:
	hf_hub_download_kwargs[key] = value
	else:
	other_kwargs[key] = value
	return hf_hub_download_kwargs, other_kwargs
	def load_adapter(self, model_id: str, adapter_name: str, is_trainable: bool = False, **kwargs: Any):
	from .mapping import PEFT_TYPE_TO_CONFIG_MAPPING
	hf_hub_download_kwargs, kwargs = self._split_kwargs(kwargs)
	if adapter_name not in self.peft_config:

This method will put the 'subfolder' keyword argument into 'hf_hub_download_kwargs'.

So here

peft/src/peft/peft_model.py

Lines 519 to 522 in 029f416

	# load weights if any
	path = os.path.join(model_id, kwargs["subfolder"]) if kwargs.get("subfolder", None) is not None else model_id
	if os.path.exists(os.path.join(path, SAFETENSORS_WEIGHTS_NAME)):

The 'subfolder' parameter is not present in kwargs anymore.

Expected behavior

please fix

[younesbelkada]

Hi @RuSignalFlag
#721 should fix the issue, thanks a lot for noticing!