perf: perf script change for qwen30b-a3b

[youngeunkwon0405]

What does this PR do ?

On-policy

• 889 --> ~1170 Token/sec/gpu
  Async 1-off
• 903 --> ~1341 Token/sec/gpu

This config change isn't the only factor of the above perf number change. It is the sum of many other changes made so far.

Issues

List issues that this PR closes (syntax):

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

[coderabbitai]

📝 Walkthrough

Walkthrough

Parallelism configuration adjustments applied to two GRPO Qwen3-30B performance recipe files: tensor model parallelism reduced from 2 to 1, sequence parallelism disabled, and pipeline parallelism increased from 1 to 2 in the async variant with vLLM tensor parallelism halved from 4 to 2.

Changes

Cohort / File(s)	Summary
GRPO Qwen3-30B performance recipes examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g-async-1off.yaml, examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g.yaml	tensor_model_parallel_size: 2 → 1; sequence_parallel: true → false; pipeline_model_parallel_size: 1 → 2 (async-1off only); vllm_cfg tensor_parallel_size: 4 → 2 (async-1off only)

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

• Review parallelism parameter consistency between files and ensure downstream references (e.g., make_sequence_length_divisible_by) remain valid with updated tensor_model_parallel_size
• Verify vLLM tensor parallelism reduction aligns with overall model partitioning strategy in async configuration

Pre-merge checks and finishing touches

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR makes major model parallelism changes (reducing tensor_model_parallel_size from 2 to 1, disabling sequence_parallel) that significantly impact performance, but provides no performance metrics, testing results, or justification.	Add performance test results comparing old and new configurations, including throughput, memory utilization, and convergence metrics with hardware/batch size context.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'perf: perf script change for qwen30b-a3b' directly addresses the main changes: performance configuration updates to Qwen 30B model scripts. It is concise and specific to the changeset.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 1

📜 Review details

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 6fc917f and af8250e.

📒 Files selected for processing (2)

• examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g-async-1off.yaml (2 hunks)
• examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g.yaml (1 hunks)

🧰 Additional context used

🧠 Learnings (2)

📓 Common learnings

Learnt from: CR
Repo: NVIDIA-NeMo/RL PR: 0
File: coderabbit-custom-pre-merge-checks-unique-id-file-non-traceable-F7F2B60C-1728-4C9A-8889-4F2235E186CA.txt:0-0
Timestamp: 2025-09-20T14:59:08.052Z
Learning: If a change could affect performance, include before-and-after performance numbers in the PR description, along with configuration and context.

Learnt from: adil-a
Repo: NVIDIA-NeMo/RL PR: 1440
File: examples/configs/sft_automodel.yaml:48-58
Timestamp: 2025-10-30T20:50:44.126Z
Learning: In DTensor configurations for MoE (Mixture of Experts) models, expert_parallel_size and data_parallel_size can be applied together without multiplying the GPU requirements. Expert Parallelism (EP) only applies to MoE layers, while Data Parallelism/FSDP applies to non-MoE layers. Therefore, configurations like expert_parallel_size: 8 and data_parallel_size: 8 are valid on an 8-GPU cluster for MoE models.

📚 Learning: 2025-10-30T20:50:44.126Z

Learnt from: adil-a
Repo: NVIDIA-NeMo/RL PR: 1440
File: examples/configs/sft_automodel.yaml:48-58
Timestamp: 2025-10-30T20:50:44.126Z
Learning: In DTensor configurations for MoE (Mixture of Experts) models, expert_parallel_size and data_parallel_size can be applied together without multiplying the GPU requirements. Expert Parallelism (EP) only applies to MoE layers, while Data Parallelism/FSDP applies to non-MoE layers. Therefore, configurations like expert_parallel_size: 8 and data_parallel_size: 8 are valid on an 8-GPU cluster for MoE models.

Applied to files:

• examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g.yaml
• examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g-async-1off.yaml

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (3)

• GitHub Check: Lint check
• GitHub Check: Post submodule check comment / Comment on PR
• GitHub Check: Post automodel integration comment / Comment on PR

🔇 Additional comments (1)

examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g-async-1off.yaml (1)

13-16: Incorrect characterization of configuration changes.

The review misidentifies the actual configuration modifications:

Parameter	Base	Async	Review claimed
tensor_model_parallel_size	1	1	2→1 ❌
pipeline_model_parallel_size	1	2	1→2 ✓
sequence_parallel	false	false	true→false ❌
vllm tensor_parallel_size	4	2	4→2 ✓

Only 2 of 4 claimed changes are accurate. The actual modifications are:

• pipeline_model_parallel_size: 1 → 2
• vllm_cfg.tensor_parallel_size: 4 → 2
• Added async_engine: true and gpu_memory_utilization: 0.8

Since the specific changes cited in the review do not match the actual code, the analysis is factually incorrect.

Likely an incorrect or invalid review comment.

[youngeunkwon0405]

Hi @guyueh1, I updated the perf recipe and also updated the numbers in the perf tracker reflecting this change.

[youngeunkwon0405]

Hi @terrykong, can we merge this recipe update PR?