cp: Dsv3 Recipe Update (2152) into r0.3.0

[ko3n1g]

beep boop [🤖]: Hi @dingqingy-nv 👋,

we've cherry picked #2152 into  for you! 🚀

Please review and approve this cherry pick by your convenience!

Summary by CodeRabbit

• New Features

  • Added public aliases for DeepSeek V3 pretraining configurations supporting multiple precision formats (BF16, FP8 variants).

• Chores

  • Updated pretraining configurations with optimized parallelism and layout settings for improved performance.
  • Added environment configuration support for FP8 mixed precision on DeepSeek models.

[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.

[ko3n1g]

/ok to test f972a84

[coderabbitai]

📝 Walkthrough

Walkthrough

Updated DeepSeek pretraining configurations for GB200 and GB300 hardware variants by propagating layout parameters, adjusting GB300_V1 parallelism settings (pipeline and virtual pipeline sizes, expert model parallel reduction), micro-batching, and recompute modules. Added public aliases for configuration variants and environment variable handling for DeepSeek fp8_mx compute dtype.

Changes

Cohort / File(s)	Summary
DeepSeek Config Updates scripts/performance/configs/deepseek/deepseek_llm_pretrain.py, scripts/performance/configs/deepseek/deepseek_workload_base_configs.py	Updated GB200 layout parameter from None to base_cfg.pp_layout. Revised GB300_V1 pretraining config with adjusted pipeline/virtual pipeline sizes (2/8), micro_batch_size (2), pp_layout, expert_model_parallel_size (64→32), and recompute_modules (["moe_act"]→["mla_up_proj"]). Added public aliases for GB300 variants (BF16_V1, FP8_CS_V1, FP8_MX_V1, NVFP4_V1/V2).
Environment Setup scripts/performance/perf_plugins.py	Added conditional branch to set del_cudnn_ln=False for DeepSeek model family with fp8_mx compute dtype in _set_model_specific_environment_variables.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Possibly related PRs

• DSV3 NVFP4 recipe on GB300 #2076 — Directly related, both PRs modify DeepSeek pretrain configs to propagate pipeline-layout into GB300/GB200 pretrain_config calls.
• ci(fix): PerfPlugin for llama #2060 — Related, both PRs modify _set_model_specific_environment_variables logic in perf_plugins.py for model-specific configuration handling.

Suggested labels

r0.3.0

Suggested reviewers

• erhoo82

🚥 Pre-merge checks | ✅ 3 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR modifies critical DeepSeek training configuration parameters (pipeline parallelism, expert parallelism, recompute modules) with high code review effort, but lacks test results, performance metrics, or convergence validation.	Include test results demonstrating no convergence regression and performance metrics for the configuration changes, or reference testing from original PR #2152.

✅ 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 indicates this is a cherry-pick of PR #2152 into branch r0.3.0, which directly corresponds to the changes shown in the raw summary (Deepseek V3 recipe updates to configuration files).
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ 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 cherry-pick-2152-r0.3.0

---

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.}