Skip to content

[rollout][trtllm] chore: adopt SleepConfig for fine-grained sleeping#5527

Closed
hchings wants to merge 3 commits intoverl-project:mainfrom
joyang-nv:hchings/gb200-setup-235B
Closed

[rollout][trtllm] chore: adopt SleepConfig for fine-grained sleeping#5527
hchings wants to merge 3 commits intoverl-project:mainfrom
joyang-nv:hchings/gb200-setup-235B

Conversation

@hchings
Copy link
Copy Markdown
Collaborator

@hchings hchings commented Mar 9, 2026

What does this PR do?

WIP on config tuning. Might move config to verl-recipe.
Reset of changes is to adopt Yuan's NVIDIA/TensorRT-LLM#11889 to avoid CPU OOM on 8 nodes for a 235B.

This MR should be merged after NVIDIA/TensorRT-LLM#11889 and #5444.

Checklist Before Starting

  • Search for similar PRs. Paste at least one query link here: ...
  • Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
    • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
    • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
    • {type} is in feat, fix, refactor, chore, test
    • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
    • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

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

Code Review

This pull request introduces changes for 235B MoE models and their configuration, primarily focusing on trtllm rollout. A new run script for qwen3-235b is added, and there are updates to the trtllm async server and rollout logic. My review focuses on ensuring the new configuration script is clear and maintainable. I've pointed out some redundant and conflicting settings in the new shell script that should be addressed to avoid confusion.

Comment on lines +102 to +103
actor_rollout_ref.rollout.name=vllm \
actor_rollout_ref.rollout.enforce_eager=True \
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.

high

There are redundant and conflicting settings for actor_rollout_ref.rollout in this script.

  • actor_rollout_ref.rollout.name is set to vllm on line 102, but overridden to trtllm on line 153.
  • actor_rollout_ref.rollout.enforce_eager is set to True on line 103, but overridden to False on line 149.

This makes the script confusing and prone to errors if someone modifies only the first occurrence. To improve clarity and maintainability, please remove these initial, overridden settings.

@hchings hchings force-pushed the hchings/gb200-setup-235B branch 2 times, most recently from 3ff2ec5 to ceac890 Compare March 12, 2026 01:50
@hchings hchings force-pushed the hchings/gb200-setup-235B branch from ceac890 to 1885689 Compare April 1, 2026 00:40
@hchings hchings changed the title [rollout][trtllm] a few changes for 235B MoE & config [rollout][trtllm] chore: adopt SleepConfig for fine-grained sleeping Apr 1, 2026
@hchings hchings closed this Apr 2, 2026
@hchings
Copy link
Copy Markdown
Collaborator Author

hchings commented Apr 2, 2026

Closed for #5841

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant