[megatron] fix: set model to eval during compute_log_prob/compute_values

[HollowMan6]

What does this PR do?

Current code doesn't seem to set model to eval when we compute_log_prob, ref_compute_log_prob, and compute_values for the megatron backend, which might go wrong if the model contains dropout layers for e.g. LoRA.

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, 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).)

[gemini-code-assist]

Code Review

This pull request correctly adds logic to set the model to evaluation mode for compute_log_prob and compute_values. My feedback focuses on improving the robustness of this change by using try...finally blocks. This ensures that the model's original training state is restored even in the event of an exception, preventing potential silent bugs during training.

[Copilot]

Pull request overview

This PR fixes a bug where models were not being set to evaluation mode during inference operations (compute_log_prob and compute_values), which could cause incorrect behavior when models contain dropout layers (e.g., in LoRA adapters).

Key changes:

• Added eval mode setting with proper mode restoration in MegatronPPOActor.compute_log_prob
• Added eval mode setting with proper mode restoration in MegatronPPOCritic.compute_values

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
verl/workers/actor/megatron_actor.py	Saves training mode state, sets modules to eval before computing log probabilities, and restores the original training mode after computation
verl/workers/critic/megatron_critic.py	Saves training mode state, sets modules to eval before computing values, and restores the original training mode after computation

The changes look correct and follow a consistent pattern. The ref_compute_log_prob mentioned in the PR description is automatically covered since it delegates to MegatronPPOActor.compute_log_prob which now has the fix.

---

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