[recipe] feat: @register_policy_loss("geo_mean"); Geometric-Mean Policy Optimization

[MzeroMiko]

What does this PR do?

This is the official implementaion of paper Geometric-Mean Policy Optimization.

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

The code has trained for 100 iterations, and is still running.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

A new policy loss function has been added into "verl/trainer/ppo/core_algos.py"

@register_policy_loss("geo_mean")
def compute_policy_loss_geo_mean(
    old_log_prob: torch.Tensor,
    log_prob: torch.Tensor,
    advantages: torch.Tensor,
    response_mask: torch.Tensor,
    loss_agg_mode: str = "token-mean",
    config: Optional[DictConfig | AlgoConfig] = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch. Tensor]:
    ...

We also added directory "examples/gmpo_trainer" for quick start.

Design & Code Changes

see above

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 introduces Geometric-Mean Policy Optimization (GMPO) by adding a new policy loss function. The implementation is a good step towards incorporating this new algorithm. However, I've identified a critical issue related to potential division-by-zero errors that could cause training instability, and another high-severity issue concerning redundant code that impacts maintainability. Addressing these points will make the implementation more robust and cleaner.

[CLAassistant]

All committers have signed the CLA.

[vermouth1992]

Please fix pre-commit according to readme and let's merge it