Updated qlora.py to fix freezing of embedding layers #217
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Outdated suggestions cannot be applied.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.
Updated the way embeddings are frozen by changing the order of operations. In the original codebase, the model was loaded, LoRA-fied and then the tokenizer was resized. The resizing resets the gradients of the embedding layers and puts them to default (i.e.
True
). This is not what you want and to fix this, you can simple do a different order: First load the base model, then resize tokenizer, then doprepare_for_kbit_training
to freeze the model's original weights, then LoRA-fy and then train.