fix(vllm-omni): retire per-request AR seed patch - #474
Open
leviking98z-rgb wants to merge 1 commit into
Open
leviking98z-rgb wants to merge 1 commit into
leviking98z-rgb wants to merge 1 commit into
Conversation
leviking98z-rgb
requested review from
celve and
zzhuoxin1508
as code owners
September 16, 2026 07:44
9 tasks
2 tasks
This branch has not been deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Remove the vLLM-Omni per-request AR seed monkey patch and its patch-table entry.
The patch replaced every unset
SamplingParams.seedwithos.urandom, based onthe assumption that reusing one
SamplingParamsobject collapses a GRPO group toidentical tokens. vLLM 0.27 and 0.28 instead keep
seed=Noneon the ordinaryrandom-sampling path, where the GPU RNG advances across batch rows. The override
is therefore unnecessary for UniRL's current single-replica recipes and makes
rollouts impossible to reproduce from the configured engine seed.
Related Issue
Refs #473.
Test Plan
SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure— passed.python -m compileall -q unirl/rollout/engine/vllm_omni/patches— passed.vllm==0.28.0,vllm-omni==0.28.0,torch==2.13.0+cu130on one NVIDIA H20: 64seed=Nonerows produced 55distinct tokens, and resetting the Torch/CUDA seed reproduced the complete
batch. No model checkpoint or dataset is involved in this sampler-level check.
v0.27.0:seed=Nonelikewise avoids a per-requestgenerator and uses the advancing batched GPU RNG.
Compatibility / Risk
Current UniRL vLLM-Omni recipes use one engine replica, so request diversity is
preserved by vLLM's advancing RNG. Independently initialized multi-replica
engines can still begin with correlated RNG streams; if that topology is added,
it should use deterministic seeds derived from request identity rather than host
entropy from
os.urandom.No config, checkpoint, data-format, or API migration is required.
Reviewer Notes
Open PR #413 removes the separate capture-flush direct registration as part of
the vLLM-Omni 0.28 migration, but retains this seed patch. No open PR was found
that removes or replaces the seed override.
AI assistance was used. The full two-file diff and all reported validation
output were reviewed before publication.
Checklist