Is it possible to "unload" the PEFT LoRA weights after mutating the base model with PeftModel.from_pretrained ?

[lxe]

Is it possible to "unload" the PEFT LoRA weights after mutating the base model with PeftModel.from_pretrained?

I'd like to load multiple LoRA models on top of a base model, and unloading the whole base model every time is time consuming. Was wondering if there's a way to un-load the PEFT model and have the base model remain in memory.

[dukesun99]

You may try model.disable_adapter_layers() or set merge_weights=false in peft_config. After disable_adapter_layers(), according to the code, if the original weights were modified, after a forward pass it will be reverted. Or if you put merge_weights=false the weights will not be modified.

Disclaimer: I am new to this lib, just based on reading the code.

[pacman100]

Hello @lxe , recent PR #227 should enable this, could you please try the main branch and let us know if that solves this issue?

[accountForIssues]

@pacman100 Does the PR allow one to swap LoRA adapters as well ? I would like to be able to load and keep a single base model in memory and swap between several LoRA trained adapters on-the-fly. (GPT2 if that matters)

I tested the main branch as of a few days ago and I wasn't able to. Had some error when trying to load another adapter after using disable_adapter_layers(). I'm new to the library so maybe I'm doing something wrong.

[pacman100]

Hello @accountForIssues, I'm currently working on it, will ping here when the PR is ready

[totaltube]

Very awaited feature!

[github-actions[bot]]

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.

[lxe]

This is now fixed in the multi adapter loading! Closing.

[Opdoop]

Hi, @lxe. I'm new to peft. The model.merge_and_unload() method looks like magic to me. Could you give a quick introduction about model.merge_and_unload()？

In my view, LoRA adds new trainable parameters/layers and inserts these layers into the base model, that is the LoRA model has additional structures on top of the base model. And we can save the merge_and_unload model and reload it with base_model.from_pretrained(unloaded_model_path) interface. But where are the additional layers and parameters?