Fixes the bug with not being able to handle loras targeting different sets of modulestosave

[saeid93]

Fixes #2422

[BenjaminBossan]

Thanks for the bug report and for providing a fix. I think this is a regression introduced by #2376, since your example works when I checkout PEFT v0.14.0.

Before proceeding, could you please create a unit test to catch the bug? It could be based on your original example and be placed here. Just note that you should not re-use the same base model for get_peft_model but create a new one each time, since the base model is mutated, which can have unintended side effects.

[githubnemo]

Hey @saeid93, thanks for your fix. I spent some time working on that issue. Good catch on the disabled adapter! I think that I've found a fix that doesn't limit the fix to ModulesToSaveWrapper and also supports other AuxiliaryTrainingWrapper such as TrainableTokensWrapper. It'd be great if you take a look at #2430 as well :)

[saeid93]

Hi @BenjaminBossan and @githubnemo

Thank you both for your feedback.
I'll add the test and check #2340 when I get a chance. Just a note about your comment this is partially related to the recent changes by #2376 but even in older versions although it doesn't return an error it returns the wrong response as the LoRAs get disabled when iterated on a non-existant moduletosave at

peft/src/peft/utils/other.py

Line 781 in 8edaae9

module.enable_adapters(False)

. Even in previous version I had to access the module to save layer manually to enable the lora like below to make it enable it in my maincode:

            model.classifiers.enable_adapters(True)

I just added a new line to ensure the lora is always enabled as @githubnemo also mentioned.

[githubnemo]

Hi @BenjaminBossan and @githubnemo

Thank you both for your feedback. I'll add the test and check #2340 when I get a chance.

Before doing that, maybe take a look at PR #2430 first to avoid doing more work than necessary, I think it fixes both the issues you mentioned.

[githubnemo]

We're going ahead with #2430 which supersedes this, so I'm closing this PR. Thank you again for this PR :)

If you have any additional comments or think that we've missed something, do not hesitate to say so.