Fix SFTtraining for new trl #2647
Merged
+4
−4
Conversation
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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Outdated suggestions cannot be applied.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.
Latest TRL redefines DataCollatorForLanguageModeling and uses that internally. We actually import it, but that override the name since we later import the same name from transformers.
https://github.com/huggingface/trl/blob/v0.18.0/trl/trainer/sft_trainer.py#L75
vs
https://github.com/huggingface/transformers/blob/v4.52.3/src/transformers/data/data_collator.py#L764
Since the signature changes unsloth sfttrainer fails becuase we initialize the transformers version with arguments expected for the sft_trainer version. My fix imports the transformers version as a different name so they don't clobber each other.
I've successfully verified that the fix works for SFT and DPO.
Failed original runs:
llama: https://colab.research.google.com/drive/1UGMaG2NQMho1z9i7ZDC9J_HYMW0S3L9k?usp=sharing
qwen: https://colab.research.google.com/drive/1w83mQOzSL6T8sbBA2GjJO8n1ycA6kzoj?usp=sharing
fixed:
qwen: https://colab.research.google.com/drive/1MmZJFz6VSHKnyPknUxDa-TDjKI1hd-1U?usp=sharing
dpo: https://colab.research.google.com/drive/1psXyPg9RFJIPUPjoeHQrUDdJA5aTT1XX?usp=sharing
Please note: With new trl 0.18.0 non completion tokens are masked by default if prompt is a valid key in the dataset