fix: fix make_sequence_length_divisible_by in config

[yuki-97]

#2053 makes megatron respect to policy.make_sequence_length_divisible_by, but some configs are missing to update this param. This PR fixes it.

Summary by CodeRabbit

Chores

• Updated sequence length alignment configurations in multiple training recipe files to optimize model training performance across different hardware configurations and parallelization strategies. Changes include increased alignment divisor values in several configurations and new dynamic calculation expressions that incorporate tensor and context parallelization factors to ensure proper sequence handling.

[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

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 163dce56-0f61-42b0-928e-25c7ee12e1e9

📥 Commits

Reviewing files that changed from the base of the PR and between 2c1e5e0 and 91b561b.

📒 Files selected for processing (8)

• examples/configs/recipes/llm/grpo-dapomath17k-dsv3-32n4g-megatron.yaml
• examples/configs/recipes/llm/grpo-dapomath17k-dsv3-megatron.yaml
• examples/configs/recipes/llm/grpo-gptoss-20b-8n4g-megatron.yaml
• examples/configs/recipes/llm/grpo-gptoss-20b-8n8g-megatron.yaml
• examples/configs/recipes/llm/grpo-nano-v2-12b-1n8g-megatron.yaml
• examples/configs/recipes/llm/performance/dapo-deepseek-v3-64n8g.yaml
• examples/configs/recipes/llm/performance/grpo-qwen3-235b-16n8g.yaml
• examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g-40K.yaml

---

📝 Walkthrough

Walkthrough

Multiple LLM training recipe configuration files were updated to adjust make_sequence_length_divisible_by settings across Megatron and DTensor policies. Changes include direct numeric value updates, expression modifications incorporating multiplied parallelism factors, and additions of new divisibility constraints. No control-flow logic or model behavior was altered.

Changes

Cohort / File(s)	Summary
Megatron Divisibility Updates examples/configs/recipes/llm/grpo-dapomath17k-dsv3-32n4g-megatron.yaml, examples/configs/recipes/llm/performance/grpo-qwen3-235b-16n8g.yaml	Updated policy.megatron_cfg.make_sequence_length_divisible_by from 4 → 16 and 1 → 8 respectively.
New Megatron Divisibility Fields examples/configs/recipes/llm/grpo-gptoss-20b-8n4g-megatron.yaml, examples/configs/recipes/llm/grpo-gptoss-20b-8n8g-megatron.yaml	Added policy.megatron_cfg.make_sequence_length_divisible_by configuration with values 2 and 4 respectively.
DTensor Divisibility Updates examples/configs/recipes/llm/grpo-nano-v2-12b-1n8g-megatron.yaml	Updated policy.dtensor_cfg.make_sequence_length_divisible_by from 1 → 8.
Complex Expression Updates examples/configs/recipes/llm/grpo-dapomath17k-dsv3-megatron.yaml, examples/configs/recipes/llm/performance/grpo-qwen3-30ba3b-4n8g-40K.yaml	Modified make_sequence_length_divisible_by expressions to multiply tensor model parallel size by 2 * context_parallel_size.
Megatron Reference Update examples/configs/recipes/llm/performance/dapo-deepseek-v3-64n8g.yaml	Switched policy.make_sequence_length_divisible_by divisor calculation from DTensor settings to Megatron settings while maintaining multiplication by 2 * context_parallel_size.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 3 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR modifies sequence length divisibility constraints across 8 configuration files without providing test results or regression validation documenting impact on model training/inference behavior.	Include test results or regression validation demonstrating the configuration changes do not negatively impact training convergence, numerical outputs, or performance metrics.

✅ 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 clearly and specifically describes the main change: fixing the make_sequence_length_divisible_by configuration parameter across multiple YAML config files.
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 unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch yukih/fix-make_sequence_length_divisible_by

---

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.}

Tip

CodeRabbit can use your project's `ruff` configuration to improve the quality of Python code reviews.

Add a Ruff configuration file to your project to customize how CodeRabbit runs ruff.

[yuki-97]

/ok to test 91b561b