Fix LoRA & (IA)³ implementation for Bart & MBart #518
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
hSterz
added a commit
to hSterz/adapter-transformers
that referenced
this pull request
Aug 5, 2023
Fix resume_from_checkpoint (adapter-hub#514) add initialization of variable so invalid checkpoints throw a understandable error Fix LoRA & (IA)³ implementation for Bart & MBart (adapter-hub#518) Fixes a critical issue in the LoRA & (IA)³ implementation of Bart & MBart, where LoRA & (IA)³ weights were not added to the intermediate and output linear layers of the model's decoder blocks. I.e., adapter configs having intermediate_lora=True or output_lora=True are added incorrectly to (M)Bart models. For LoRA, this does not affect the default config, for (IA)³ it does (intermediate_lora=True). To ensure correct addition of weights in the future, get_adapter() tests are updated to count the number of modules added per adapter. Fix python3.7 Compatibility (adapter-hub#510) Compatibility with python3.8+ and pytorch1.12.1+ Restore compatibility in GPT-2 LoRALinear bias init (adapter-hub#525) Fix compacter init weights (adapter-hub#516) Update doc chapter "Getting Started" (adapter-hub#527) Update version to 3.2.1 Fix Notebook01 Dataset column_rename (adapter-hub#543) Update doc chapter "Adapter Methods" (adapter-hub#535) Do not stale issues labeled as bugs (adapter-hub#550)
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Outdated suggestions cannot be applied.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.
This PR fixes a critical issue in the LoRA & (IA)³ implementation of Bart & MBart, where LoRA & (IA)³ weights were not added to the intermediate and output linear layers of the model's decoder blocks.
I.e., adapter configs having
intermediate_lora=Trueoroutput_lora=Trueare added incorrectly to (M)Bart models. For LoRA, this does not affect the default config, for (IA)³ it does (intermediate_lora=True).To ensure correct addition of weights in the future,
get_adapter()tests are updated to count the number of modules added per adapter.