Fix default Lora/ (IA)^3 scaling in forward #770
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
4 tasks
Open
julian-fong
added a commit
to julian-fong/adapters
that referenced
this pull request
Dec 24, 2024
TimoImhof
approved these changes
Dec 29, 2024
Co-authored-by: TimoImhof <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Resolves issue described in #760.
IMPORTANT: this fix restores weights compatibility with adapter-transformers. Compatibility to previous adapters versions is kept via a compat patch.
Details
The current implementation of LoRA/ (IA)^3 in
adaptersversions < 1.1.0 does not correctly implement adapter states scaling via the LoRAalphaattribute, effectively ignoringalphaand always applying a scaling of 1.0.This PR restores the correct original behavior as found in adapter-transformers/ original LoRA implementation.
As this change breaks all adapters pre-trained using
adaptersversions 0.1.0 - 1.0.1, a backward compatibility patch is introduced that automatically setsalpha = rfor LoRAs for adapters that were trained using affected versions. This ensures all previous adapters continue to behave exactly as trained (ie give the exact same output using newer versions).