Log compaction-aware rollout metrics - #2710
Merged
Merged
Conversation
Shi-Dong
requested review from
fzyzcjy,
maocheng23 and
yueming-yuan
as code owners
August 22, 2026 05:54
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Claude Code Review
This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.
Tip: disable this comment in your organization's Code Review settings.
Shi-Dong
force-pushed
the
shi/260821-rollout-step-metrics
branch
from
August 22, 2026 05:58
519fdd9 to
3912a50
Compare
guapisolo
approved these changes
Aug 22, 2026
yueming-yuan
approved these changes
Aug 25, 2026
arpitg1991
added a commit
to arpitg1991/miles
that referenced
this pull request
Sep 8, 2026
…sample metrics build_dp_schedule needs a micro-batch count that is a multiple of dp_size * mb_group. It splits multi-sample bins but not singletons, so an odd number of rows over max_tokens_per_gpu asserts under --arena-train-segments all. The upstream Terminus 2 recipe shares this gap. Pad the batch with zero-loss sibling segments of the shortest kept row: same rollout_id and reward, all-zero loss mask, status COMPLETED, so the episode advantage and loss weight are unchanged. Drop rollout/compaction_segments_mean. miles log_rollout_data already logs rollout/num_training_samples and rollout/episode_raw_reward (radixark#2710). ADR-0011 records the precedent and the pad.
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
3 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
rollout/num_training_samplesfor the finalized samples in each rollout steprollout/episode_raw_rewardby averaging samples within each original rollout, then averaging rollouts equallyWhy
Session compaction can turn one rollout into many training samples. The existing raw-reward metric gives those rollouts proportionally more weight, and W&B does not currently expose how many samples a step trains on. These metrics make both effects visible without changing the meaning of existing metric histories.
Testing
python -m pytest --confcutdir=tests/fast/ray/rollout tests/fast/ray/rollout/test_metrics.py -q(23 passed)uvx black --check miles/ray/rollout/metrics.py tests/fast/ray/rollout/test_metrics.pypython -m ruff check miles/ray/rollout/metrics.py tests/fast/ray/rollout/test_metrics.pygit diff --check