search

huggingface / peft

🤗 PEFT: State-of-the-art Parameter-Efficient Fine-Tuning.
https://huggingface.co/docs/peft
Apache License 2.0
15.04k stars 1.44k forks source link

Add Special Optimizer for LoRA training #1803

Closed fangzhaozhang closed 14 hours ago

fangzhaozhang commented 1 month ago

Feature request

Our research group has studied speeding up LoRA training with new optimizer features, and we would like to contribute to integrating our proposed method in current huggingface libraries. The innovation is that we introduce a special gradient preconditioner which has been verified to enhance LoRA training quality significantly via extensive experiments. Our paper can be accessed at https://arxiv.org/pdf/2402.02347 and it's recently accepted by ICML 2024.

We haven't contributed to HF libraries before and am not sure whether it's possible to integrate our method to existing code base, if so, what are the steps we should follow?

Motivation

Our feature request is not related to a problem. We empirically verified that a new optimizer variant helps LoRA training significantly and we feel integrating it to existing HF trainer setting would benefit potential users who want to try it out.

Your contribution

We are happy to implement all the code and merge it to existing HF code base, though we are not familiar with steps required.

BenjaminBossan commented 1 month ago

Thanks for suggesting to add this new optimization method to PEFT. From a quick glance at the paper, I think PEFT would be a good place to add it.

If I understand this right, the new method would be about creating a new PyTorch optimizer class that works especially well with LoRA. At the moment, we don't have optimizers in PEFT (or training code, for that matter), but there is a PR to add LoRA+ (#1509), so you could check that PR out for inspiration (like module structure).

If you already have a repo with code, feel free to share it. Otherwise, just open a PR. You can create a draft PR and share your work even if not yet finished to get some early feedback.

fangzhaozhang commented 1 month ago

Thanks for reply. We have opened a draft PR (#1807) and pushed some initial code for our method & simple test following LoRA+ implementation(#1509).

Any feedback would be appreciated.

github-actions[bot] commented 2 weeks ago

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.