[megatron] fix: add megatron checkpoint patch

[Begunner]

What does this PR do?

Add a monkey patch to force to release the memory of input tensors and their grads in CheckpointFunction.

Unreleased GPU memory occurs during RL or SFT training of MoE models (e.g., Qwen3-30B-A3B/Qwen3-VL-30B-A3B) using image verlai/verl:vllm012.dev3 with checkpoint enabled (recompute_method=uniform, recompute_granularity=full, recompute_num_layers=1). Key observations:

• No residual memory with dense models (e.g., Qwen3-VL-2B/Qwen3-8B) or without checkpoint.
• Residual memory size correlates positively with recompute_num_layers (halved when set to 2).

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

This pull request introduces a monkey patch to address a GPU memory leak in Megatron's checkpointing mechanism when used with Mixture-of-Experts (MoE) models. The patch manually releases memory for input tensors and their gradients within the CheckpointFunction.backward method. The approach is sound for fixing the leak. However, the patch is applied unconditionally, which could pose a risk to non-MoE models. I've suggested making its application conditional based on whether an MoE model is in use.

[gemini-code-assist]

The comment "Apply checkpoint patch for MoE models" indicates this patch is specific to Mixture-of-Experts models. However, it's being applied unconditionally. This could introduce risks or unintended side effects for non-MoE models that also use checkpointing. It would be safer to apply this patch conditionally, only when an MoE model is detected. You can check for this using self.engine_config.expert_model_parallel_size > 1.

Suggested change

	# Apply checkpoint patch for MoE models
	from verl.models.mcore.patch import apply_patch_checkpoint
	apply_patch_checkpoint()
	# Apply checkpoint patch for MoE models
	if self.engine_config.expert_model_parallel_size > 1:
	from verl.models.mcore.patch import apply_patch_checkpoint
	apply_patch_checkpoint()

[ISEEKYAN]

megatron PR link NVIDIA/Megatron-LM#3267