Update sft trainer to include better packing #3100
Closed
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
Ishan-Kumar2
changed the title
qgallouedec
reviewed
Apr 2, 2025
Member
qgallouedec
left a comment
qgallouedec
left a comment
There was a problem hiding this comment.
Thanks for your contribution, and sorry for the late reply. I've made some comments. Can you please also:
- include a test for
pack_example_smater - modify the packing argument in
SFTConfigto allow for this packing method
trl/data_utils.py
Outdated
| return knapsacks, [[tup[1]] for tup in numbers] | ||
|
|
||
|
|
||
| def pack_examples_smarter(examples: dict[str, list[list]], seq_length: int) -> dict[str, list[list]]: |
Member
There was a problem hiding this comment.
wdyt of
Suggested change
| def pack_examples_smarter(examples: dict[str, list[list]], seq_length: int) -> dict[str, list[list]]: | |
| def pack_examples_knapsack(examples: dict[str, list[list]], seq_length: int) -> dict[str, list[list]]: |
or even
Suggested change
| def pack_examples_smarter(examples: dict[str, list[list]], seq_length: int) -> dict[str, list[list]]: | |
| def pack_examples_knapsack_greedy(examples: dict[str, list[list]], seq_length: int) -> dict[str, list[list]]: |
trl/trainer/sft_trainer.py
Outdated
Comment on lines
364
to
367
| packing: bool, | ||
| formatting_func: Optional[Callable[[dict], str]], | ||
| dataset_name: str, | ||
| pack_smart: bool = False, |
Member
There was a problem hiding this comment.
maybe this is better:
Suggested change
| packing: bool, | |
| formatting_func: Optional[Callable[[dict], str]], | |
| dataset_name: str, | |
| pack_smart: bool = False, | |
| packing: Union[bool, str], | |
| formatting_func: Optional[Callable[[dict], str]], | |
| dataset_name: str, |
and use if packing == "knapsack":
Contributor
Author
|
Hi @qgallouedec sorry for the delay, I had my final exams :) |
Member
|
I think we've arrived at something very satisfactory with FFD see #3521, so I'm closing this PR. Thanks for all your hard work! |
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.
What does this PR do?
Fixes #2466
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines.
Who can review?
@qgallouedec @shirinyamani @AIR-hl
Hi everyone, I am new to TRL, have attempted a fix at the issue 2466 here.
The greedy knapsack tries to fit as many examples as possible without overshooting, the remaining ones are chunked as before by breaking it up into smaller example size (<= max_seq_len). Also I havent made this the default packing and kept it under a flag.
Please let me know if this is close to what was expected, I can add tests then, thanks!