fix: stabilize GB200 16n4g issue-2579 repro#3013
Open
jinglinglingling wants to merge 3 commits into
Open
Conversation
Capture the CUDA13/vLLM0.20 compatibility fixes that let the 16n4g Qwen235 repro advance stably, including Triton MoE routing, safer policy/refit lifecycle handling, and optional lazy/combined weight-sync paths needed for this environment. Signed-off-by: Linglin Jing <linglinj@nvidia.com>
jinglinglingling
added
the
CI:Lfast
Contributor
Author
|
/ok to test 0946a24 |
Run ruff formatting adjustments on the GB200 issue-2579 Python changes so lint and CI quality checks can pass without altering runtime behavior. Signed-off-by: Linglin Jing <linglinj@nvidia.com>
Merge latest main to satisfy branch freshness checks and keep the GB200 issue-2579 fixes aligned with current upstream behavior. Signed-off-by: Linglin Jing <linglinj@nvidia.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Issue
related to #2579.
closes #2812
Summary
Validated that the GB200
grpo-qwen3-235b-16n4grepro is now stable with CUDA13 + vLLM0.20 compatibility fixes.In the latest long run (
max_num_steps=30), training advanced through Step 9 and entered Step 10.The run stopped due to Slurm time limit, not due to runtime crash.
Changes
weights+kv_cache) path to reduce refit/wakeup instability.tensordict) for worker environments.Validation
grpo-qwen3-235b-16n4grun progressed stably past historical Step-4 failure point.ActorDiedError,RayChannelTimeoutError, or NCCL collective timeout observed in this run.