[megatron] fix: LoRA with activation checkpointing

[HollowMan6]

What does this PR do?

Resolves #4063 (comment)

Refer to NVIDIA-NeMo/Megatron-Bridge#1750

This PR enable grad on TransformerBlock inputs when only adapters are trainable, so that we can still have gradients when activation checkpointing is enabled.

• With PP>1, received tensors from other stages have requires_grad=True, so checkpoint backward is always called.
• With PP=1 and frozen base model, embedding outputs have requires_grad=False. This means
  CheckpointFunction.backward() is never called, and LoRA gradients inside the checkpoint are never computed.

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, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • 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).)

_{✨ Presented to you with Mind Lab - A Lab for Experiential Intelligence.}

[gemini-code-assist]

Code Review

This pull request introduces a fix for using LoRA with activation checkpointing in Megatron when the base model is frozen. The core of the solution is a new utility function, maybe_enable_recompute_inputs_grad, which monkey-patches TransformerBlock.forward to ensure that input tensors require gradients. This is crucial for triggering the backward pass of checkpointed blocks. This utility is then integrated into the actor and critic workers. The overall approach is sound and effectively addresses the issue. I have one suggestion to improve code correctness related to a type hint.

[HollowMan6]

Convergence tests:

[HollowMan6]

Waiting for NVIDIA-NeMo/Megatron-Bridge#1762