[algo] fix: normalize GDPO advantages over responses

[lucky9-cyou]

What does this PR do?

This PR updates GDPO's final batch-level advantage normalization to operate over generated responses instead of valid response tokens.
In the current GDPO implementation, compute_grpo_outcome_advantage returns outcome-level advantages with token shape, i.e. one scalar advantage is broadcast to all valid response tokens. The final GDPO normalization then applies:

advantages = verl_F.masked_whiten(new_advantage, response_mask) * response_mask

This computes whitening statistics over all valid tokens. As a result, longer responses contribute more copies of the same outcome-level advantage to the batch mean/std.

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, fully_async, one_step_off
  • 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).)
• 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 modifies the advantage calculation in compute_gdpo_outcome_advantage to compute and whiten advantages at the response level rather than the token level. The reviewer identified a critical issue where whitening a single valid sequence (or a batch size of 1) can cause a division-by-zero crash, and provided a code suggestion to safely handle this edge case.

[gemini-code-assist]

If the batch size is 1, or if only one sequence in the batch has a valid response (i.e., response_level_mask.sum() is 1), verl_F.masked_whiten will call masked_var which raises a ValueError: The sum of the mask is one, which can cause a division by zero. This will crash the training or validation loop.

To prevent this, we should check if response_level_mask.sum() > 1 before applying masked_whiten. If there is at most one valid sequence, we can safely set the whitened advantages to zero. Additionally, converting response_level_mask to the same dtype as response_level_advantage ensures compatibility across different PyTorch versions and hardware backends.

Suggested change

	response_level_advantage = verl_F.masked_mean(new_advantage, response_mask, axis=-1)
	response_level_mask = response_mask.sum(dim=-1) > 0
	response_level_advantage = verl_F.masked_whiten(response_level_advantage, response_level_mask)
	response_level_advantage = verl_F.masked_mean(new_advantage, response_mask, axis=-1)
	response_level_mask = (response_mask.sum(dim=-1) > 0).to(response_level_advantage.dtype)
	if response_level_mask.sum() > 1:
	response_level_advantage = verl_F.masked_whiten(response_level_advantage, response_level_mask)
	else:
	response_level_advantage = torch.zeros_like(response_level_advantage)