feat(sc): stream PPO actor and critic minibatches - #3963
Open
tianyi-zhang-02 wants to merge 5 commits into
Open
Conversation
Signed-off-by: Tianyi Zhang <123608656+tianyi-zhang-02@users.noreply.github.com>
Signed-off-by: Tianyi Zhang <123608656+tianyi-zhang-02@users.noreply.github.com>
Signed-off-by: Tianyi Zhang <123608656+tianyi-zhang-02@users.noreply.github.com>
github-actions
Bot
added
the
Documentation
tianyi-zhang-02
marked this pull request as ready for review
September 2, 2026 15:11
…inibatches # Conflicts: # docs/guides/single-controller.md
svcnvidia-nemo-ci
added
the
waiting-on-maintainers
…inibatches # Conflicts: # nemo_rl/algorithms/single_controller.py
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What does this PR do?
Lets Single Controller PPO stream actor and Megatron critic minibatches from native TQ replay instead of materializing one merged training payload.
The critic now has an explicit begin/train-chunk/finish/abort lifecycle. Chunks accumulate gradients and correctly normalized diagnostics, then take one optimizer step at the end of each critic epoch. The controller retains replay rows across critic and actor epochs, cleans up partial failures, skips invalid chunks, and rejects drop budgets that cannot preserve data-parallel batch divisibility.
Relates to #2625.
Validation
Current head
884e238a1fdd2f84cb44a3deb2042981e91d05c8, merged with upstreammainate831622435180c88d6adcc9fce46bfd88a85f3e5.The main refresh composes the current token-capture finalizer cleanup for canonical and staging rows with PPO's multi-epoch row retention. The current-head CPU-capable suite covers that merged control flow, including failure cleanup.
The GPU runs were collected at pre-refresh head
2a9b108a38a0be424c9693c806cd0e41dd5985b5. The refreshed head was revalidated with the full CPU-capable suite above; GPU-only Megatron split parity remains skipped in the current environment.Functional entry point:
Before your PR is ready for review