patch sft_trainer to favor max_seq_length over max_length in config #2669
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Outdated suggestions cannot be applied.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.
newest trl has changed from max_seq_length to max_length. In order to maintain old behavior unsloth will favor max_seq_length if it's in the args or if it's set inside the model for sft trainer only.
tested with qwen3:
trl==0.15.2 takes 26 mins
https://colab.research.google.com/drive/1A9DJ6SYsgYWtDciL2wvw-A-1ja238gs1?usp=sharing
upgrade to trl==0.18.1 changes the run to take 13 mins.
https://colab.research.google.com/drive/1IYGn6psJyfhZz-Zabg0IrrhUOBnRMtpR?usp=sharing
Issue is that max_length is defaulted to 1024 and unsloth prepare sft dataset takes this as preference. I added a check to the trl patches for sft_trainer file, and add explicit handling of max_length and max_seq_length. Now the notebook takes the expected ~20ish minutes.
https://colab.research.google.com/drive/1A6uj-VZsRPvPBLNy0ySPyC1kyV07Xeud?usp=sharing
You can also inspect the compiled files to see that it's only sfttrainer that gets impacted.