Garbage collect on every train / ref model step #209
Merged
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
Summary: Test Plan: Reviewers: Subscribers: Tasks: Tags:
meta-cla
bot
added
the
CLA Signed
allenwang28
reviewed
Sep 22, 2025
allenwang28
approved these changes
Sep 22, 2025
Ritesh1905
reviewed
Sep 22, 2025
src/forge/actors/trainer.py
Outdated
There was a problem hiding this comment.
question: Why do we do CP save on evert step?
There was a problem hiding this comment.
This is a Titan checkpointer impl detail. What actually happens is that it checks if it should save, which is determined by the checkpoint frequency attr found in our config. If it shouldn't checkpoint it just returns. See here.
A much much much better name would be maybe_save IMO
There was a problem hiding this comment.
Curious why isn't it a "finally" step that's done by default?
Ritesh1905
reviewed
Sep 22, 2025
src/forge/actors/reference_model.py
Outdated
There was a problem hiding this comment.
Can you explain what is this doing internally?
There was a problem hiding this comment.
photomz
pushed a commit
to photomz/forge
that referenced
this pull request
Oct 25, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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Outdated suggestions cannot be applied.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 is this PR?
As reported in #201, there appeared to be a memory leak in the components using TorchTitan - RLTrainer and ReferenceModel. This PR manually calls the TorchTitan garbage collection util on every step, which eliminates the memory leak.
How was this PR tested?
With
gc_freq: 100000(never running essentially)Notice the yellow line (reference model) and the ref line (trainer model) consistently going up.
With
gc_freq: 1(run every step)Notice the yellow line (reference model) and red line (trainer model) stays relatively flat.
FAQs
The theory is that Monarch doesn't know that it is able to free memory after returning from an Endpoint. So every time the Endpoint (eitherActually Titan disables garbage collection manually to improve performance. We have to re-enable it with this PR.forwardortrain_step) is called, new memory is allocated. This theory definitely requires further investigation, but the fix lends credibility.