Repository navigation
verl buffer trainer: background releases, per-round phase timing and hybrid MFU accounting - #327
Merged
Merged
Conversation
2 tasks done
michaellchung
force-pushed
the
sparse-weight-update
branch
from
November 2, 2026 17:31
1bc0eaf to
fc2d311
Compare
michaellchung
force-pushed
the
background-release-mfu
branch
2 times, most recently
from
November 2, 2026 17:42
d564b00 to
ec4c078
Compare
michaellchung
force-pushed
the
sparse-weight-update
branch
from
November 2, 2026 17:42
fc2d311 to
6665b65
Compare
michaellchung
force-pushed
the
background-release-mfu
branch
from
November 2, 2026 18:30
ec4c078 to
fa54896
Compare
michaellchung
force-pushed
the
sparse-weight-update
branch
2 times, most recently
from
November 2, 2026 18:42
5211599 to
c27856a
Compare
michaellchung
force-pushed
the
background-release-mfu
branch
2 times, most recently
from
November 2, 2026 19:27
91226de to
7ed7c17
Compare
michaellchung
force-pushed
the
sparse-weight-update
branch
from
November 2, 2026 19:27
c27856a to
10f5fed
Compare
michaellchung
force-pushed
the
background-release-mfu
branch
2 times, most recently
from
November 2, 2026 20:09
c60a94d to
933f436
Compare
michaellchung
force-pushed
the
sparse-weight-update
branch
from
November 3, 2026 02:05
42c8fa0 to
fe46eba
Compare
michaellchung
force-pushed
the
background-release-mfu
branch
from
November 3, 2026 02:05
933f436 to
7ec663b
Compare
…hybrid MFU accounting - --rl-release-background: the staged release is written on the training thread, then encoded, digested and uploaded on a worker thread while the next rounds train; at most one release in flight, its error raises at the next publication; the release log stays in version order and the release origin is computed on the training thread; the publisher is closed when the run raises. - rl_round_phases: per-round phase timing emitted on the trainer islands only; no clock reads for phase accounting while it is off. - hybrid_flops: Ascend MFU accounting and a CLI that prints step/wall MFU per trained round from a trainer event tape. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
michaellchung
force-pushed
the
background-release-mfu
branch
from
November 3, 2026 03:29
7ec663b to
6ef9f6e
Compare
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
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
--rl-release-background(buffer trainer, needs--rl-release-delta-full-every N): the staged release is written on the training thread, then encoded, digested and uploaded on a worker thread while the next rounds train. At most one release is in flight; the next publication waits for it and re-raises its error. The release log stays in version order, the release origin is computed on the training thread, and the publisher is closed when the run raises.rl_round_phases: per-round phase timing, emitted on trainer islands only; no clock reads for phase accounting while it is off.yeto.rl.hybrid_flops: Ascend MFU accounting and a CLI that printsstep_mfu/wall_mfuper trained round from a trainer event tape, for hybrid models (Nemotron-H) whose framework MFU counter does not cover them.docs/MILES_RL.md.Behavior changes
None by default: background releases are opt-in; single trainer only.
Test plan
pytest tests/test_rl_hybrid_params.py tests/test_rl_engine_driver.py tests/test_verl_sample_role.py tests/test_rl_verl_buffer_trainer.py tests/test_rl_sparse_weight_update.py: 172 passed.scripts/precommit_check.pyon the changed files: no errors.Stacked on #326.
🤖 Generated with Claude Code