Add support for modules_to_save in FastModel.get_peft_model

[l1ghtsource]

Problem Description

Currently, the FastModel.get_peft_model method does not support the modules_to_save argument.

If a user passes modules_to_save=["score"] when training an LLM classifier, the head (score) remains a regular Linear module and is not wrapped in ModulesToSaveWrapper.

This results in the head’s weights not being saved in adapter_model.safetensors along with the LoRA adapters, making reuse or further fine-tuning impossible.

Solution

• Added support to pass the modules_to_save argument into the internal LoRA configuration.
• The head (or any specified modules) is now correctly wrapped in ModulesToSaveWrapper.

[danielhanchen]

Thanks - I'll also refactor to allow all other items!