feat: add dapo recipe and test

[ZhiyuLi-Nvidia]

What does this PR do ?

Verified the test and recipe: https://wandb.ai/nvidia/nemo-rl/runs/ks1cpy6o/logs?nw=nwuserzhiyul
Update test and recipe: https://wandb.ai/nvidia/nemo-rl/runs/g5r4xwvf
Update test and recipe with 1.5k seq length: https://wandb.ai/nvidia/nemo-rl/runs/eobc39eb

Issues

List issues that this PR closes (syntax):

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

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 run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

Summary by CodeRabbit

• New Features

  • Added comprehensive performance configuration and test suite for DAPO DeepSeek v3 with 64-node setup, including GRPO training parameters, resource allocation, data configuration, and automated metrics validation.

• Chores

  • Updated internal submodule dependencies.

_{✏️ Tip: You can customize this high-level summary in your review settings.}

[github-actions]

✅ Submodule Fast-Forward Check Results

Check based on commit: bbd6933 (PR #1617 from zhiyul/dapo_sv3_pr)

✅ Submodules that are properly updated:

Megatron-LM: ✅ PR branch is ahead of main branch (fast-forward)

All submodule changes look good! ✨

[coderabbitai]

📝 Walkthrough

Walkthrough

Updates submodule tracking for 3rdparty/Megatron-LM to use a new branch variant with yarn fix support. Adds new DAPO DeepSeek v3 performance configuration file and corresponding test script with environment setup, execution, and metrics evaluation logic.

Changes

Cohort / File(s)	Summary
Submodule Updates .gitmodules, 3rdparty/Megatron-LM-workspace/Megatron-LM	Updated Megatron-LM submodule branch reference from yuya/nemo-rl-use-dev to yuya/nemo-rl-use-dev-w-yarn-fix and bumped commit pointer to 1d646249...
DAPO DeepSeek v3 Config & Tests examples/configs/recipes/llm/performance/dapo-deepseek-v3-64n8g.yaml, tests/test_suites/llm/performance/dapo-deepseek-v3-64n8g.sh	Added comprehensive YAML configuration for DAPO DeepSeek v3 performance with grpo settings, policy/data/cluster parameters, and logging integrations. Added Bash test script setting up environment, executing experiment via run_grpo_math.py, converting TensorBoard logs to JSON, and conditionally evaluating metrics against thresholds.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~25 minutes

• YAML configuration verification: Validate all grpo, policy, cluster, and logging settings align with intended performance tuning parameters and are compatible with the underlying infrastructure
• Test script logic: Verify environment variable handling (NVLS disable, checkpoint override), parameter passing to Python scripts, and conditional metric evaluation logic
• Submodule branch correctness: Confirm the new branch variant (...w-yarn-fix) is the intended tracking target and compatible with current codebase

Pre-merge checks and finishing touches

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR introduces major changes (DAPO recipe and performance tests) but PR description is in draft status with unchecked pre-review items and no documented test results, performance benchmarks, or regression testing information.	Run test suite locally and document results in PR description; provide before-and-after performance comparisons; validate numerics stability; complete all pre-review checklist items; update PR with comprehensive testing information.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Title check	✅ Passed	The title clearly summarizes the main changes: adding a new DAPO (DeepSeek v3) recipe configuration and corresponding test script.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.

✨ 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 zhiyul/dapo_sv3_pr

---

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.}

[github-actions]

✅ Submodule Fast-Forward Check Results

Check based on commit: 679273f (PR #1617 from zhiyul/dapo_sv3_pr)

✅ Submodules that are properly updated:

Megatron-LM: ✅ PR branch is ahead of main branch (fast-forward)

All submodule changes look good! ✨

[ZhiyuLi-Nvidia]

@guyueh1 could you take a review?

[guyueh1]

@ZhiyuLi-Nvidia would you add this test to the tests/test_suites/performance_h100.txt?

[ZhiyuLi-Nvidia]

@ZhiyuLi-Nvidia would you add this test to the tests/test_suites/performance_h100.txt?

Thank you @guyueh1, just added.

[terrykong]

@ZhiyuLi-Nvidia is this ready?

@guyueh1 to review

[ZhiyuLi-Nvidia]

@ZhiyuLi-Nvidia is this ready?

Yeap.

@guyueh1 I have address you comment to "add this test to the tests/test_suites/performance_h100.txt". Could you take another look?