[megatron] fix: patch support newer mcore version

[HollowMan6]

What does this PR do?

Patch support newer mcore version

Tested on NVIDIA/Megatron-LM@bbbedbb

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[gemini-code-assist]

Code Review

The pull request successfully updates the Megatron-Core (mcore) patching logic to support version 0.13.0 and newer. The changes correctly handle the updated return signatures of internal methods like get_query_key_value_tensors and _adjust_key_value_for_inference, and appropriately skip legacy patches that are no longer required in newer mcore versions. The implementation is robust and maintains backward compatibility with mcore 0.12. No high-severity issues were identified in the changes.

[Copilot]

Pull request overview

This PR enhances Megatron Core version compatibility by updating the Multi-Latent Attention (MLA) patch to support mcore >= 0.13. The key change is to use defensive unpacking for the return values of get_query_key_value_tensors, and to conditionally apply the legacy patch only for mcore < 0.13, as newer versions have the fix upstream.

Changes:

• Changed from direct tuple unpacking to slice-based unpacking (qkv[:3]) to handle variable return values across mcore versions
• Added conditional patch application for get_query_key_value_tensors - only applies to mcore < 0.13
• Added explanatory comments about the version-specific behavior

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[vermouth1992]

What's the meaning of this change?

[HollowMan6]

In NVIDIA/Megatron-LM@382eeea#diff-d7e453d550e7f78221ea3445018264015c1e83c0db892a41253bde71d069a202L247, the return value for get_query_key_value_tensors has changed from query, key, value to query, key, value, q_compressed, kv_compressed. kv_compressed is not used in the forward pass so it just gets ignored, here I write in this way to take the first 3 returned values for backward compatibility.

I just found that q_compressed is actually used for DSA like DeepSeek V3.2, so I made and pushed a modification to fix it.

[vermouth1992]

Do we still need this whole patch for newer version megatron?

[HollowMan6]

Yes, as unfortunately NVIDIA/TransformerEngine#2629 hasn't been merged into TE.

[HollowMan6]

and neither NVIDIA/Megatron-LM#3003 for mcore, if one of them gets merged, then we don't need this patch to support flash attention for MLA.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[gemini-code-assist]

Code Review

This pull request aims to add support for newer versions of megatron-core by adapting to changes in the attention mechanism API. The changes correctly handle a new potential return value (q_compressed) from get_query_key_value_tensors and pass it to core_attention when the dsa attention variant is used. The conditional patching based on the mcore version is also a good approach to maintain backward compatibility. I have one suggestion to improve robustness by adding an assertion to ensure q_compressed is available when required, which will make potential configuration errors easier to debug.