Add HybridEP support for Qwen3 235B H100

[rhmukundan]

Summary by CodeRabbit

• Chores
  • Updated workload dispatcher configuration for optimized performance in specific model scenarios.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between deb49eb and 8fcf2cf.

📒 Files selected for processing (1)

• scripts/performance/configs/qwen/qwen3_workload_base_configs.py

---

📝 Walkthrough

Walkthrough

This change modifies Mixture of Experts (MoE) dispatcher configuration for two QWEN3 235B A22B pretraining configurations targeting H100 GPUs. It disables the MoE all-to-all overlap feature and explicitly specifies the hybridep dispatcher backend.

Changes

Cohort / File(s)	Summary
MoE Dispatcher Configuration scripts/performance/configs/qwen/qwen3_workload_base_configs.py	Disabled moe_a2a_overlap and added moe_flex_dispatcher_backend="hybridep" in QWEN3_235B_A22B_PRETRAIN_CONFIG_H100_BF16_V1 and QWEN3_235B_A22B_PRETRAIN_CONFIG_H100_FP8_CS_V1 configurations.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Possibly related PRs

• qwen x86 flex backend #2351: Modifies the same config entries for QWEN3 235B A22B configurations with MoE dispatcher settings.
• fix: perf configs 2 #2393: Changes MoE dispatcher configuration for QWEN3 pretraining including moe_flex_dispatcher_backend and moe_a2a_overlap settings.
• cp: qwen gbs 2x (2280) into r0.3.0 #2369: Edits the same config file and modifies moe_flex_dispatcher_backend for QWEN3 workload configurations.

Suggested reviewers

• ko3n1g
• malay-nagda
• erhoo82

🚥 Pre-merge checks | ✅ 3 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR enables HybridEP optimization for Qwen3 235B on H100 GPUs but lacks performance benchmarks, test results, or convergence validation in the PR description.	Add performance benchmark results (throughput and TFLOP/sec in BF16/FP8) and convergence validation to PR description; update performance-summary.md with new H100 HybridEP configuration results.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main change: adding HybridEP support for Qwen3 235B H100 by setting moe_flex_dispatcher_backend="hybridep" in the configuration file.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

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

✨ Finishing Touches

• 📝 Generate docstrings (stacked PR)
• 📝 Generate docstrings (commit on current branch)

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch rmukundan/qwen3_235b_hybridep_h100

Tip

Try Coding Plans. Let us write the prompt for your AI agent so you can ship faster (with fewer bugs).
Share your feedback on Discord.

---

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.}