[megatron] refactor: Refactor policy update failure handling from exception to warning in MegatronPPOActor to support FP16 training

[mirrorboat]

Refactor policy update failure handling from exception to warning in MegatronPPOActor to support FP16 training

What does this PR do?

Replaced NotImplementedError with a warning log when policy updates fail due to gradient overflow in FP16 training. This prevents abrupt termination and allows continued execution while alerting users to potential precision issues.

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

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 refactors the policy update failure handling in MegatronPPOActor to be more robust for FP16 training. Instead of raising a NotImplementedError when a policy update fails (e.g., due to gradient overflow), it now logs a warning and continues. This prevents training from crashing on transient errors. My review includes a suggestion to enhance this by adding a counter for consecutive failures to prevent silent training stalls in case of persistent issues.

[gemini-code-assist]

While replacing NotImplementedError with a warning is a good step to handle transient gradient overflows in FP16 training without crashing, it introduces a risk of silent failures. If policy updates fail repeatedly, the model will not learn, but the training script will continue to run, potentially wasting significant compute resources. This could happen if there's a more persistent instability issue.

To make this more robust, I suggest introducing a counter for consecutive update failures. If the number of consecutive failures exceeds a configurable threshold, it should raise an exception to halt the training. This provides a safety net against silent training failures while still allowing for occasional, transient issues.

You would need to:

1. Initialize self.consecutive_update_failures = 0 in the __init__ method.
2. Reset the counter with self.consecutive_update_failures = 0 inside the if update_successful: block (around line 812).
3. Implement the failure counting and check in this else block as suggested below.

Suggested change

	else:
	raise NotImplementedError
	logger.warning("Policy update failed, no update is made. This can be caused by gradient overflow in fp16 training.")
	else:
	self.consecutive_update_failures += 1
	logger.warning(
	f"Policy update failed ({self.consecutive_update_failures} consecutive), no update is made. "
	"This can be caused by gradient overflow in fp16 training."
	)
	max_failures = self.config.get("max_consecutive_update_failures", 10)
	if self.consecutive_update_failures > max_failures:
	raise RuntimeError(
	f"Exceeded max consecutive policy update failures ({max_failures}). "
	"Training is likely unstable."
	)

[mirrorboat]

@ISEEKYAN @vermouth1992 Could you please review this PR that supports FP16 training for Megatron ? Thanks.