cp: feat: default packed_sequence to True in all finetune recipes (2284) into r0.3.0

[ko3n1g]

beep boop [🤖]: Hi @yaoyu-33 👋,

we've cherry picked #2284 into  for you! 🚀

Please review and approve this cherry pick by your convenience!

Summary by CodeRabbit

Release Notes

• Configuration Changes
  • Sequence packing is now enabled by default during finetuning for most model architectures (Gemma, Qwen, Llama, GPT-OSS, Nemotron, Moonlight, and OLMoE), improving training efficiency and memory utilization.
  • GLM 4.5 continues to use sequence packing disabled by default; users can override this setting per their requirements.

[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 dca08ae

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR updates the default value of the packed_sequence parameter from False to True across multiple model recipe finetune helper functions and the default_squad_config utility. The GLM4.5 recipe adds an assertion to enforce packed_sequence as False despite the signature change. One test file adds a mock attribute.

Changes

Cohort / File(s)	Summary
Gemma and Gemma3 Recipes src/megatron/bridge/recipes/gemma/gemma2.py, src/megatron/bridge/recipes/gemma/gemma3.py	Updated packed_sequence parameter default from False to True in finetune helper functions.
GLM4.5 Recipe src/megatron/bridge/recipes/glm/glm45.py	Added packed_sequence: bool to GLM45FinetuneKwargs, changed default in _glm45_finetune_common to True, and added runtime assertion (assert not packed_sequence) to enforce False behavior. Finetune config helpers explicitly set packed_sequence to False.
GPT-OSS Recipe src/megatron/bridge/recipes/gpt_oss/gpt_oss.py	Updated packed_sequence parameter default from False to True in _gpt_oss_finetune_common.
Llama Recipe src/megatron/bridge/recipes/llama/llama3.py	Updated packed_sequence parameter default from False to True in _llama3_finetune_common.
Moonlight Recipe src/megatron/bridge/recipes/moonlight/moonlight_16b.py	Updated packed_sequence parameter default from False to True in _moonlight_finetune_common.
Nemotron Recipes src/megatron/bridge/recipes/nemotronh/nemotron_3_nano.py, src/megatron/bridge/recipes/nemotronh/nemotronh.py	Updated packed_sequence parameter default from False to True in finetune helper functions.
OLMoE Recipe src/megatron/bridge/recipes/olmoe/olmoe_7b.py	Updated packed_sequence parameter default from False to True in _olmoe_finetune_common.
Qwen Recipes src/megatron/bridge/recipes/qwen/qwen2.py, src/megatron/bridge/recipes/qwen/qwen3.py, src/megatron/bridge/recipes/qwen/qwen3_moe.py, src/megatron/bridge/recipes/qwen/qwen3_next.py	Updated packed_sequence parameter default from False to True in finetune helper functions. Qwen3-Next also explicitly sets packed_sequence: False in recommended config with documentation note.
Finetune Utilities src/megatron/bridge/recipes/utils/finetune_utils.py	Updated default_squad_config function signature to default packed_sequence parameter to True instead of False.
Test Updates tests/unit_tests/recipes/test_qwen_recipes.py	Added context_parallel_size = 1 attribute to mock model configuration used by Qwen test suite.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Possibly related PRs

• feat: default packed_sequence to True in all finetune recipes #2284 — Makes identical default value changes for packed_sequence across the same set of finetune helper functions.
• Nemotron-3-Nano Model Support #1914 — Modifies the _nemotron_3_nano_finetune_common function that was introduced by Nemotron-3-Nano support.

Suggested labels

r0.3.0

Suggested reviewers

• yaoyu-33
• cuichenx

🚥 Pre-merge checks | ✅ 3 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Test Results For Major Changes	⚠️ Warning	PR changes default packed_sequence from False to True across 13+ recipe files affecting training behavior, but lacks test results, convergence validation, or performance metrics.	Add test results, convergence validation, and performance metrics to PR documentation demonstrating no regressions across affected models.

✅ 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 describes the main change: defaulting packed_sequence to True across all finetune recipes. It is specific, concise, and directly reflects the primary purpose of the changeset.
Docstring Coverage	✅ Passed	Docstring coverage is 90.00% 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-2284-r0.3.0

---

No actionable comments were generated in the recent review. 🎉

🧹 Recent nitpick comments

src/megatron/bridge/recipes/glm/glm45.py (1)

417-417: Default packed_sequence=True immediately triggers the assertion on line 446.

The signature defaults to True, but line 446 asserts not packed_sequence. Any direct call to _glm45_finetune_common() without explicitly passing packed_sequence=False will raise AssertionError. While the public configs both override this to False, the contradictory default is confusing and fragile.

Consider defaulting to False here (unlike the other recipes) since GLM 4.5 doesn't support packing, and preferably use raise ValueError(...) instead of assert for input validation that should not be stripped by -O.

Suggested fix

-    packed_sequence: bool = True,
+    packed_sequence: bool = False,

And on line 446:

-    assert not packed_sequence, "Packed sequence is not supported for GLM 4.5 finetuning"
+    if packed_sequence:
+        raise ValueError("Packed sequence is not supported for GLM 4.5 finetuning")

Also applies to: 446-446

---

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.}

[cuichenx]

need thorough testing before merging

[ko3n1g]

merged via the finetune PR