Skip to content

Conversation

@Datta0
Copy link
Collaborator

@Datta0 Datta0 commented Oct 22, 2025

Previously we were performing sleep and wakeup in _prepare_inputs.
Now that we're doing it in _generate_and_score_completions as of #3492 , we do not need to do it in _prepare_inputs anymore.
In fact, doing so would cause double copy of memory resulting in OOM.

Caveat: This might not work for older versions of trl (0.16.0 and 0.17.0 mostly. v0.18 and above should work fine)

@danielhanchen danielhanchen merged commit bbdab30 into unslothai:main Oct 23, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants