Lora saving is broken with the new PEFT version

[myluki2000]

Describe the bug

You changed the used peft version in commit https://github.com/oobabooga/text-generation-webui/commit/64e3b44e0f5d7a008d031c3d328804b1617c8c0e to use the most up-to-date version from the peft repository. However it seems like Lora saving is borked in that version, at least on my machine. All adapter_model.bin files created are basically empty (always 443 Bytes in size). The same thing happened when I tried to update the peft version in my own fork a few days ago (for an unrelated feature I wanted to implement).

Does this happen for anyone else or is this an issue on my end?

Is there an existing issue for this?

• [X] I have searched the existing issues

Reproduction

Just train any model and look at the created file in the loras directory

Screenshot

No response

Logs

none

System Info

Windows 10
Python 3.10
RTX 3090

[oobabooga]

The update is necessary for multi-lora support. The issue with saving LoRAs has been reported by @mcmonkey4eva here

https://github.com/huggingface/peft/issues/317

A workaround for now is to install peft==0.2.0

[win10ogod]

can Second training lora?

[myluki2000]

Oops, yea. I missed checking your discussion on the PR's page. I'll leave this issue open for now anyways so other users that have this issue can find information about it.

[FaceDeer]

Yesterday I ran install.bat and then trained a Lora on a small test text, and it seemed to work fine. This issue may have been resolved by changes since it was filed.

[github-actions[bot]]

This issue has been closed due to inactivity for 6 weeks. If you believe it is still relevant, please leave a comment below. You can tag a developer in your comment.