Feature request

Currently, when using transformers.pipeline to load a PEFT fine-tuned model (e.g., "ybelkada/opt-350m-lora"), the pipeline loads only the base model without applying the LoRA adapters. This behavior is misleading because users would expect to get the fine-tuned version of the model rather than just the base model.

For example:

import transformers
pipeline = transformers.pipeline("text-generation", model="ybelkada/opt-350m-lora")

This code loads the base OPT model instead of the LoRA fine-tuned opt-350m-lora model.

I propose one of the following improvements:

1. Add a warning message when loading a PEFT fine-tuned model without applying LoRA adapters. For example:

   "Warning: You are loading a LoRA fine-tuned model, but the LoRA adapters have not been applied. Use PeftModel.from_pretrained() to correctly load the fine-tuned version."

2. Automatically detect and apply LoRA adapters when using pipeline, similar to how AutoModelForCausalLM.from_pretrained() works with PEFT models.

Motivation

This issue is problematic because it creates a false assumption that users are working with the fine-tuned model when, in reality, they are only using the base model. Many users might not realize this and get incorrect results without knowing why. A clear warning or automatic application of LoRA would significantly improve user experience and reduce confusion.

Your contribution

I can help test the implementation if needed. Let me know if there are any specific areas where contributions would be useful.