Skip to content

[Proposal] Add FRoD tuner for full-rank parameter-efficient fine-tuning #3244

@Bane-Elvin

Description

@Bane-Elvin

Feature request

I would like to propose adding FRoD as a new tuner to huggingface/peft.

FRoD (Full-Rank Efficient Fine-Tuning with Rotational Degrees for Fast Convergence) is a parameter-efficient fine-tuning method that uses a shared full-rank basis together with sparse learnable rotational updates. The goal is to preserve the expressive capacity of full-rank adaptation while keeping the number of trainable parameters small.

I believe FRoD is a good fit for PEFT because:

  • it targets standard linear layers in Transformer blocks,
  • it follows the adapter-style fine-tuning workflow used by other PEFT tuners,
  • it can be integrated through get_peft_model(...),
  • and it provides a structured full-rank alternative to low-rank adaptation methods.

Paper:

Code:

If this proposal is accepted, I plan to adapt FRoD to the latest upstream PEFT code structure and submit the implementation as a PR.

Your contribution

Yes. I can prepare and submit a PR for the core FRoD tuner.

My planned contribution includes:

  • adding PeftType.FROD,
  • adding src/peft/tuners/frod/ with config.py, layer.py, model.py, and __init__.py,
  • registering the tuner in the current PEFT registration flow,
  • adding save/load support,
  • adding tests for integration with PEFT model wrapping,
  • and adding documentation for the new method.

I will target the latest peft main branch rather than submitting an old fork directly. If preferred, I can keep the first PR focused on the core tuner and submit runnable task examples in a follow-up PR.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions