Bug with 'PromptEncoderConfig' in Newer Versions of PEFT

[nikiduki]

System Info

peft==0.15.1
transformers==4.50.2

Who can help?

@BenjaminBossan @sayakpaul

When creating a PromptEncoderConfig with task_type="SEQ_CLS" on newer versions of peft (>=0.15.0), the get_peft_model method fails with the error:
AttributeError: 'PromptEncoderConfig' object has no attribute 'modules_to_save'

When running the code on version 0.14.0, no errors occur.

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

from peft import PromptEncoderConfig, get_peft_model
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_name = "google-bert/bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
base_model = AutoModelForSequenceClassification.from_pretrained(model_name)

ptuning_config = PromptEncoderConfig(task_type="SEQ_CLS", num_virtual_tokens=20)

model = get_peft_model(base_model, ptuning_config)

Expected behavior

When running the code on peft >=0.15.0, the get_peft_model method fails with the following error:
AttributeError: 'PromptEncoderConfig' object has no attribute 'modules_to_save'

In contrast, running the code on peft 0.14.0 does not produce any errors.

[BenjaminBossan]

Thanks for the report, this is indeed a regression.

@githubnemo Before the change, we have used model.modules_to_save, now peft_config.modules_to_save. Should we go back to the former?

[githubnemo]

AFAICS this was done in ed3c828 to fix #2206 and should not be reverted. Instead, we should probably just check if the config has a modules_to_save key. Maybe #2481 will fix this.

[githubnemo]

We released v0.15.2 from a patch branch as a quickfix for this. #2481 is still open but should be ready soon as well :)