Ko3n1g/chore/reapply 2152 and 2209

[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 configurations for DeepSeek and Qwen3 models.
  • Optimized parallelism and CUDA graph settings for improved performance.
  • Simplified configuration aliases for GB300 variants.
  • Enhanced runtime environment variable handling for specific model-precision combinations.

[coderabbitai]

Caution

Review failed

The pull request is closed.

📝 Walkthrough

Walkthrough

This PR modifies DeepSeek and Qwen3 pretraining configurations to adjust parallelism settings, pipeline layouts, and CUDA graph scopes, while adding model-specific environment variable handling for DeepSeek fp8_mx compute dtype.

Changes

Cohort / File(s)	Summary
DeepSeek Layout & Configuration scripts/performance/configs/deepseek/deepseek_llm_pretrain.py, scripts/performance/configs/deepseek/deepseek_workload_base_configs.py	Updated GB200 to use base_cfg.pp_layout instead of None; redefined GB300_V1 with simplified micro_batch_size, reduced pipeline/virtual pipeline, and updated CUDA graph scope; consolidated NVFP4 variants to use V1/V2 bases directly, reducing cross-variant dependencies.
Qwen3 Expert Parallelism & CUDA Graph scripts/performance/configs/qwen/qwen3_workload_base_configs.py	Increased expert_model_parallel_size from 16 to 32 in GB300_FP8_CS_V2, removed virtual_pipeline_model_parallel_size=12, and added cuda_graph_scope with attention and MoE components; added explicit expert_model_parallel_size=32 to GB200_FP8_CS_V2.
Performance Plugin Environment Handling scripts/performance/perf_plugins.py	Added conditional branch to disable CUDNN layer normalization deletion when model family is "deepseek" and compute dtype is "fp8_mx".

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Possibly related PRs

• PR #2271 — Directly reverts the layout argument change and cudnn-norm special-case additions from this PR
• PR #2210 — Modifies identical QWEN3_235B_A22B_PRETRAIN_CONFIG_GB300_FP8_CS_V2 and GB200_FP8_CS_V2 entries with the same expert parallelism and cuda_graph_scope adjustments
• PR #2186 — Makes overlapping changes to DeepSeek configs and perf_plugins cudnn LayerNorm handling

Suggested labels

r0.3.0, Run CICD, cherry-pick

Suggested reviewers

• erhoo82
• dingqingy-nv

✨ 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/chore/reapply-2152-and-2209

---

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.}