fix: correct KL metrics in rollout importance sampling

[hjh0119]

align KL metrics with https://github.com/volcengine/verl/blob/main/verl/trainer/ppo/rollout_corr_helper.py

[gemini-code-assist]

Summary of Changes

Hello @hjh0119, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses an inaccuracy in the calculation and description of Kullback-Leibler (KL) divergence metrics within the GRPO and RLHF trainers. It corrects the direction of the KL divergence from "KL(training_policy || rollout_policy)" to "KL(rollout_policy || training_policy)", ensuring consistency with the intended metric and a reference implementation. This change impacts both the mathematical formulas presented in the documentation and the actual code implementation of the direct KL estimator.

Highlights

• KL Divergence Definition Correction: The primary intent of this pull request is to correct the definition of KL divergence used in rollout importance sampling from "KL(training_policy || rollout_policy)" to "KL(rollout_policy || training_policy)", aligning with a reference implementation and ensuring the metric accurately reflects the deviation of the rollout policy from the training policy.
• Direct KL Estimator Corrected: The calculation for the direct KL estimator ("kl") in grpo_trainer.py and rlhf_trainer/grpo_trainer.py has been correctly updated to masked_mean(-log_ratio, completion_mask) to compute "KL(π_rollout || π_training)".
• Documentation and Comment Updates: Both English and Chinese documentation files, along with code comments in the GRPO and RLHF trainers, have been updated to reflect the intended "KL(π_rollout || π_training)" definition for both direct and K3 estimators.
• K3 KL Estimator Inconsistency: While the documentation and comments for the K3 estimator ("k3_kl") now state it estimates "KL(π_rollout || π_training)", the underlying mathematical formula "E_rollout[ ρ - log ρ - 1 ]" (where "ρ = π_training / π_rollout") used in both the documentation and implementation still mathematically corresponds to estimating "KL(π_training || π_rollout)" when sampled from the rollout policy. This indicates a potential inconsistency between the stated intent and the actual K3 implementation.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request aims to correct the KL divergence metrics to align with KL(π_rollout || π_training). The direct estimator kl has been correctly updated. However, the K3 estimator k3_kl implementation remains unchanged in one file and incorrectly documented in others. It still calculates KL(π_training || π_rollout), while its corresponding comments and documentation have been updated to state it calculates KL(π_rollout || π_training). This introduces an inconsistency. I've provided suggestions to correct the k3_kl implementation and its mathematical representation in the documentation to ensure both metrics consistently estimate KL(π_rollout || π_training). A similar issue exists in swift/megatron/trainers/grpo_trainer.py which is not part of the diff but should be addressed for consistency.