[megatron] fix: resolve backward propagation error in megatron_actor due to shared logits tensor in-place modification

[HelloWorld686]

What does this PR do?

Fixes gradient computation conflict in verl/workers/actor/megatron_actor.py when entropy regularization is enabled:

• Root Cause: The entropy calculation entropy = vocab_parallel_entropy(logits) fails during backward propagation because log_probs = vocab_parallel_log_probs_from_logits(logits, label) performs in-place modifications on the logits tensor earlier in the code. This corrupts the original computation graph needed for gradient calculation.
• Fix: Decouples tensor dependencies by cloning logits before entropy calculation to preserve the original computation graph while maintaining existing log_probs computation flow.

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

1. Run modified training script:

examples/ppo_trainer/run_qwen2-7b_math_gsm8k_megatron.sh \
--actor_rollout_ref.actor.entropy_coeff=0.01

2. The following error is observed (before repair):

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.

[gemini-code-assist]

Code Review

This PR fixes a critical bug in the backward propagation step when entropy regularization is enabled, by cloning the logits tensor. I've added a high-severity comment suggesting an optimization to make the tensor clone conditional, preventing unnecessary memory allocation and computation when entropy is not calculated.

[gemini-code-assist]

The logits.clone() operation is performed unconditionally. This clone is only necessary when calculate_entropy is True to prevent in-place modification from corrupting the computation graph. When calculate_entropy is False, this clone is redundant and introduces unnecessary overhead. To optimize, perform the clone operation conditionally.

Suggested change

	logits_bak = logits.clone()
	if calculate_entropy:
	entropy = vocab_parallel_entropy(logits)
	ret["entropy"] = entropy
	log_probs = vocab_parallel_log_probs_from_logits(logits, label)
	log_probs = vocab_parallel_log_probs_from_logits(logits_bak, label)
	log_probs = log_probs.masked_fill(~label_mask, 0.0)
	if calculate_entropy:
	logits_bak = logits.clone()
	entropy = vocab_parallel_entropy(logits_bak)
	ret["entropy"] = entropy
	log_probs = vocab_parallel_log_probs_from_logits(logits_bak, label)
	else:
	log_probs = vocab_parallel_log_probs_from_logits(logits, label)

[ETOgaosion]

Thanks for contribution, @HelloWorld686 Could you consider this advice?

[HelloWorld686]

Good idea, thanks

[ETOgaosion]

Notice that here

verl/verl/utils/megatron/tensor_parallel.py

Line 133 in d9a6a31

vocab_parallel_logits.sub_(sum_softmax_times_logits)

And

verl/verl/utils/megatron/tensor_parallel.py

Lines 136 to 137 in d9a6a31

	# recover vocab_parallel_logits
	vocab_parallel_logits.add_(sum_softmax_times_logits)

vocab_parallel_logits shall remain the same, so current logic seems OK?

[HelloWorld686]

Notice that here

verl/verl/utils/megatron/tensor_parallel.py

Line 133 in d9a6a31

vocab_parallel_logits.sub_(sum_softmax_times_logits)

And

verl/verl/utils/megatron/tensor_parallel.py

Lines 136 to 137 in d9a6a31

	# recover vocab_parallel_logits
	vocab_parallel_logits.add_(sum_softmax_times_logits)

vocab_parallel_logits shall remain the same, so current logic seems OK?

During the forward pass of vocab_parallel_entropy, it does not alter the values of the logits; however, it saves the logits for use in backpropagation. In contrast, vocab_parallel_log_probs_from_logits modifies the logits in-place, which can cause errors during the backpropagation step of vocab_parallel_entropy.

[vermouth1992]

In megatron, entropy is only computed for observation, not for regularization. logits.clone consumes too much memory and is not affordable. We need a kernel to make entropy regularization happen

[ETOgaosion]

#2210 has been merged as mcore side fused kernel function.

Shall we fix the correctness here?

[HelloWorld686]

#2210 has been merged as mcore side fused kernel function.

Shall we fix the correctness here?

Thank you for your feedback. While memory efficiency is important, ensuring functional correctness should be our top priority—especially since the use_fused_kernels path doesn’t cover all use cases. The clone operation prevents bugs in the non-use_fused_kernels path. (We can’t just force-enable fused kernels to avoid bugs, can we?)

What do you think about prioritizing correctness across all scenarios first, then optimizing memory afterward?

[HelloWorld686]

@ETOgaosion , I've rebased my PR with the latest code and now some of the tests have passed, including the one that previously failed. However, the PR run was cancelled a few days ago and there hasn't been any further activity. Could you please re-review the PR when you have a chance? Thank you!