DSv3 EP=8 for B200, PP8-VP2 for B300 BF16, Lm3.1 405B TP4-CP1 GB300 FP8-CS

[malay-nagda]

What does this PR do ?

• Change EP=16 to EP=8 for B200 (both BF16 and FP8)
• Change PP16-VP1 to PP8-VP2 for B300 BF16
• Change TP2-CP2 to TP4-CP1 for Llama3.1 405B GB300 FP8-CS

Changelog

BASE_DEEPSEEK_V3_CONFIG,
    ...
    - expert_model_parallel_size=16,
    + expert_model_parallel_size=8,

DEEPSEEK_V3_PRETRAIN_CONFIG_B300_BF16_V2 = replace(
    DEEPSEEK_V3_PRETRAIN_CONFIG_B300_V2,
    pipeline_model_parallel_size=8,
    virtual_pipeline_model_parallel_size=2,
)

LLAMA31_405B_PRETRAIN_CONFIG_GB300_FP8_CS_V2 = replace(
    LLAMA31_405B_PRETRAIN_CONFIG_GB300_FP8_CS_V1,
    tensor_model_parallel_size=4,
    pipeline_model_parallel_size=8,
    context_parallel_size=1,

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
  • Adjusted performance workload configuration parameters for optimization.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[coderabbitai]

📝 Walkthrough

Walkthrough

The expert_model_parallel_size parameter in the DEEPSEEK_V3_PRETRAIN_CONFIG_B200_V1 configuration is reduced from 16 to 8. No other configuration values, control flow, or logic are affected by this change.

Changes

Cohort / File(s)	Summary
Configuration Parameter Update scripts/performance/configs/deepseek/deepseek_workload_base_configs.py	Reduced expert_model_parallel_size from 16 to 8 in DEEPSEEK_V3_PRETRAIN_CONFIG_B200_V1.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

🚥 Pre-merge checks | ✅ 2 | ❌ 2

❌ Failed checks (1 warning, 1 inconclusive)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR changes DeepSeek V3 expert_model_parallel_size from 16 to 8 on B200 without required performance benchmarks, convergence data, or testing confirmation per CONTRIBUTING.md guidelines.	Add performance metrics (throughput, GPU utilization, memory efficiency), convergence validation, explanation of EP=8 optimization for B200, and testing confirmation.
Title check	❓ Inconclusive	The title is highly technical and contains multiple unrelated configuration changes, making it difficult to identify the primary change. While it mentions 'DSv3 EP=8 for B200' which relates to the code change, the title also includes unrelated items (PP8-VP2 for B300, Lm3.1 405B, etc.) that are not reflected in the actual changeset.	Simplify the title to focus on the primary change: 'Change DeepSeek V3 expert parallelism to 8 for B200' or similar. Remove unrelated configuration items not present in this PR's changeset.

✅ Passed checks (2 passed)

Check name	Status	Explanation
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.

_{✏️ 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 malay/b200_dsv3_ep8

---

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.}