Skip to content

feat(sc): abort stale in-flight rollouts early under over_sampling - #3263

Open
RayenTian wants to merge 5 commits into
mainfrom
ruit/sc-entrypoint-gate
Open

feat(sc): abort stale in-flight rollouts early under over_sampling#3263
RayenTian wants to merge 5 commits into
mainfrom
ruit/sc-entrypoint-gate

Conversation

@RayenTian

@RayenTian RayenTian commented Jul 17, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Support aborting stale in-flight prompt groups when they fall outside the sampler selection window.
  • Add W&B metrics for evicted stale prompt groups (train/evicted_stale_prompt_groups) and aborted stale in-flight groups (train/aborted_stale_inflight_groups).
  • Fix the legacy GRPO tokenizer decode error by explicitly disabling tokenization-space cleanup.

Tests

Recipe: examples/configs/recipes/llm/performance/grpo-llama3.1-8b-instruct-2n8g-async-1off.yaml

disable kv cache recompute

image

enable kv cache recompute

image

@copy-pr-bot

copy-pr-bot Bot commented Jul 17, 2026

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@RayenTian
RayenTian changed the base branch from yukih/sc-entrypoint to main July 31, 2026 04:29
@RayenTian
RayenTian force-pushed the ruit/sc-entrypoint-gate branch 3 times, most recently from 5aaedbb to cc02515 Compare August 1, 2026 06:17
@RayenTian RayenTian added the CI:Lfast Runs a fast test suite and re-use nightly `main` container (but sync dependencies to PRs version) label Aug 1, 2026
@RayenTian

Copy link
Copy Markdown
Contributor Author

/ok to test cc02515

Comment thread nemo_rl/experience/rollouts.py Outdated
@RayenTian
RayenTian marked this pull request as ready for review August 1, 2026 07:05
@RayenTian
RayenTian requested review from a team as code owners August 1, 2026 07:05
@RayenTian
RayenTian requested review from mehraakash and yuki-97 August 1, 2026 07:05
@RayenTian
RayenTian force-pushed the ruit/sc-entrypoint-gate branch from cc02515 to 728de3f Compare August 1, 2026 13:58
@RayenTian

Copy link
Copy Markdown
Contributor Author

/ok to test 728de3f

@RayenTian

Copy link
Copy Markdown
Contributor Author

/ok to test 2019c54

@yuki-97 yuki-97 left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

thanks for supporting this, changes overall LGTM. left some comments and one q for nemo-gym path.

Comment thread nemo_rl/models/generation/vllm/vllm_worker_async.py Outdated
Comment thread tests/unit/experience/test_rollouts.py
Comment thread tests/unit/single_controller/test_train_pump.py
Comment thread nemo_rl/experience/rollouts.py Outdated
Comment thread nemo_rl/algorithms/single_controller.py
) # Default 15 minutes

# Propagate cancellation to the Ray worker and its vLLM request.
try:

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

curious have you tested whether nemo-gym path works? since the changes in vLLM all seem to be native async path.

@RayenTian
RayenTian force-pushed the ruit/sc-entrypoint-gate branch from 2019c54 to e5ed92b Compare August 4, 2026 05:48
Signed-off-by: ruit <ruit@nvidia.com>
Signed-off-by: ruit <ruit@nvidia.com>
Signed-off-by: ruit <ruit@nvidia.com>
Signed-off-by: ruit <ruit@nvidia.com>
Signed-off-by: ruit <ruit@nvidia.com>
@RayenTian
RayenTian force-pushed the ruit/sc-entrypoint-gate branch from 0416763 to 43baf36 Compare August 4, 2026 06:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CI:Lfast Runs a fast test suite and re-use nightly `main` container (but sync dependencies to PRs version)

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants