fix: perf configs 2

[ko3n1g]

What does this PR do ?

Add a one line overview of what this PR aims to accomplish.

Changelog

• Add specific line by line info of high level changes in this PR.

GitHub Actions CI

See the CI sectionin the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.

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 add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to # (issue)

Summary by CodeRabbit

• Chores
  • Updated pretraining configuration settings to enable token dispatch improvements across multiple model variants.

[coderabbitai]

📝 Walkthrough

Walkthrough

This pull request adds token-level MOE dispatcher configuration to multiple Qwen3 pretraining model configurations. Specifically, cfg.model.moe_token_dispatcher_type is set to "flex" across several configuration functions (gb200, gb300, b300, b200, h100, and variants), complementing existing MOE dispatcher backend settings.

Changes

Cohort / File(s)	Summary
Qwen3 MOE Dispatcher Configuration scripts/performance/configs/qwen/qwen3_llm_pretrain.py	Adds token-level MOE dispatcher type assignment ("flex") to multiple Qwen3 pretraining configuration functions alongside existing dispatcher backend settings.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

Possibly related PRs

• qwen x86 flex backend #2351 — Sets moe_flex_dispatcher_backend configuration in QWEN3 configs, related dispatcher backend setup.
• cp: qwen gbs 2x (2280) into r0.3.0 #2369 — Modifies QWEN3 MOE dispatcher configuration, related to dispatcher type settings.

Suggested labels

performance

Suggested reviewers

• yaoyu-33
• erhoo82

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (1 warning, 1 inconclusive)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR adds MOE token dispatcher type configuration changes affecting performance-critical behavior, but PR description lacks test results, performance metrics, or regression analysis.	Update PR description with before-and-after performance metrics and convergence data for affected Qwen3 configurations to validate no regressions.
Title check	❓ Inconclusive	The title 'fix: perf configs 2' is vague and generic, lacking specificity about which performance configurations are being fixed.	Consider a more descriptive title such as 'fix: enable flex token dispatcher for Qwen3 pretrain configs' to clearly communicate the primary change.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
Merge Conflict Detection	✅ Passed	✅ No merge conflicts detected when merging into main

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch ko3n1g/fix/perf-configs-2

---

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]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

scripts/performance/configs/qwen/qwen3_llm_pretrain.py (1)

78-84: ⚠️ Potential issue | 🟠 Major

moe_token_dispatcher_type is not set in the gb300 config for qwen3_235b, unlike all other sibling configs.

Every other qwen3_235b_a22b_pretrain_config_* and qwen3_30b_a3b_pretrain_config_* function explicitly sets cfg.model.moe_token_dispatcher_type (to either "flex" or "alltoall"). This function sets moe_flex_dispatcher_backend but omits the dispatcher type, so it will silently use whatever default the framework provides.

If this is intentional, a comment explaining the omission would help. Otherwise, it likely needs "flex" to match the gb200 variant:

Proposed fix

     cfg.model.moe_flex_dispatcher_backend = base_cfg.moe_flex_dispatcher_backend
+    cfg.model.moe_token_dispatcher_type = "flex"
 
     set_qwen3_common_configs(cfg)

As per coding guidelines: "Do not add arbitrary defaults for configs, be as explicit as possible."