[megatron] fix: revert megatron actor refactor

[vermouth1992]

What does this PR do?

• Revert megatron actor changes in this PR that causes perf degradation: [megatron] refactor: refactor MegatronPPOActor #3206
• We have to revert following PRs that modify the files too: [megatron] Add TIS support to megatron backend #3513 and [worker] fix: get all multi_modal_inputs keys with in a microbatch #3315
• We will add them back when we fix the problem

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).)

[gemini-code-assist]

Code Review

This pull request reverts several previous changes in verl/workers/actor/megatron_actor.py to address a performance degradation. The changes involve significant refactoring of loss computation and data handling logic. My review identified a critical issue that could cause an IndexError when processing empty multi-modal inputs. Additionally, I found two high-severity issues where logging and printing are no longer restricted to the main process, which would lead to cluttered logs. I have provided code suggestions to resolve these problems.

[gemini-code-assist]

Accessing batch["multi_modal_inputs"][0] without checking if the list is empty can lead to an IndexError. The current check if "multi_modal_inputs" in batch: only verifies the key's existence, not whether the list contains any elements. An empty list will cause a crash.

Suggested change

	if "multi_modal_inputs" in batch:
	if "multi_modal_inputs" in batch and batch["multi_modal_inputs"]:

[gemini-code-assist]

The removal of the if torch.distributed.get_rank() == 0: guard will cause print(config) to execute on every process, which can flood the logs with redundant information. It's best practice to restrict such print statements to a single rank (usually rank 0) to maintain clean and readable logs.

Suggested change

	print(config)
	if torch.distributed.get_rank() == 0:
	print(config)

[gemini-code-assist]

The logger.warning_once call is no longer guarded by if torch.distributed.get_rank() == 0:. While warning_once ensures the message is logged only once per process, removing the rank check will cause it to fire on every process, leading to cluttered output. This type of warning should typically be limited to rank 0.

Suggested change

	logger.warning_once(
	"For memory-efficient computation, enable fused kernels via "
	"`actor_rollout_ref.model.use_fused_kernels=True`. "
	"The current `clone()` operation ensures correctness but increases memory usage."
	)
	if torch.distributed.get_rank() == 0:
	logger.warning_once(
	"For memory-efficient computation, enable fused kernels via "
	"`actor_rollout_ref.model.use_fused_kernels=True`. "
	"The current `clone()` operation ensures correctness but increases memory usage."
	)

[Altair-Alpha]

@vermouth1992 Hi, we're updating our verl code recently and found similar problem. It turns out that in MegatronPPOActor, the following line
config.finalize_model_grads_func = finalize_model_grads
should not be removed in __init__, otherwise the value of grad_norm would be abnormaly high along with lower reward. Other changes works good. Hope this piece of information may help you :)

[HollowMan6]

Hi @vermouth1992! I noticed that this PR reverts the megatron actor changes related to #3315. #3315 should be independent of #3216 (either with or without should still be fine). Is there any specific reason to get that part in #3315 reverted? Without those changes in #3315, the error when we train the mixture modal dataset will remain unresolved, so it would be a good idea to add them back. I just opened PR #3641 for adding it back, and it has been tested on my side.

File "verl/workers/actor/megatron_actor.py", line 639, in update_policy
  metric_micro_batch = self.forward_backward_batch(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "verl/workers/actor/megatron_actor.py", line 587, in forward_backward_batch
  losses_reduced = forward_backward_func(
                   ^^^^^^^^^^^^^^^^^^^^^^
File "megatron/core/pipeline_parallel/schedules.py", line 595, in forward_backward_no_pipelining
  output_tensor, num_tokens = forward_step(
                              ^^^^^^^^^^^^^
File "megatron/core/pipeline_parallel/schedules.py", line 402, in forward_step
  output_tensor, loss_func = forward_step_func(data_iterator, model)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "verl/workers/actor/megatron_actor.py", line 497, in forward_step
  multi_modal_inputs[key] = torch.cat(
                            ^^^^^^^^^^
RuntimeError: torch.cat(): expected a non-empty list of Tensors