feat(sc): overlap NeMo-Gym spinup with deferred vLLM load - #3499
Open
yuki-97 wants to merge 7 commits into
Open
Conversation
Signed-off-by: Yuki Huang <yukih@nvidia.com>
Signed-off-by: Yuki Huang <yukih@nvidia.com>
Signed-off-by: Yuki Huang <yukih@nvidia.com>
yuki-97
added
the
CI:Lfast
Contributor
Author
|
/ok to test aa6b9ee |
Contributor
Author
|
/ok to test 4800540 |
…nit-time key Signed-off-by: Yuki Huang <yukih@nvidia.com>
yuki-97
force-pushed
the
yukih/sc-parallel-init
branch
from
August 5, 2026 14:40
4800540 to
87d655f
Compare
Contributor
Author
|
/ok to test 87d655f |
Signed-off-by: Yuki Huang <yukih@nvidia.com>
yuki-97
marked this pull request as ready for review
August 5, 2026 17:02
Contributor
Author
|
/ok to test a50365b |
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
defer_model_load=True) so NeMo-Gym spins up in parallel with the vLLM weight load and trainer build.SetupTimingMetricsdataclass +print_setup_timing_summaryintonemo_rl.algorithms.metric_utils, shared by SC and grpo.worker_setup_time_sand dropsparallel_wall_time_s/parallel_init_enabled(grpo-only). Everything else matches grpo.py.grpo-llama3.1-8b-instruct-2n8g-async-1off-single-controller-streaming2's threshold since that value was tested by non-streaming instead of streaming 2 by mistake in feat(sc): setup + entrypoint #3266. w/ the fix in fix: fix single controller grad_norm #3282,mean(data["train/grad_norm"], 2, 0)is0.07307618069979879; w/o the fix, it's0.04454974799106518, so update to> 0.06.Test plan
uv run pytest -q tests/unit/single_controller/grpo-llama3.1-8b-instruct-2n8g-async-1off-single-controller-streaming2grpo-qwen2.5-math-1.5b-instruct-1n8g-megatron-single-controller-synctests/functional/grpo_async_gym_single_controller.sh