Add report for gsm8k alignment experiment.

[chenyushuo]

Description

As the title says.

Checklist

Please check the following items before code is ready to be reviewed.

• Code has passed all tests
• Docstrings have been added/updated in Google Style
• Documentation has been updated
• Code is ready for review

[gemini-code-assist]

Summary of Changes

Hello @chenyushuo, 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 introduces a detailed report documenting a thorough alignment experiment between the veRL and Trinity machine learning frameworks, specifically focusing on their performance on the GSM8K mathematical reasoning benchmark. The report outlines the necessary code adjustments made to both systems to facilitate controlled comparisons across multiple experimental setups, ultimately validating their functional equivalence when configured identically and explaining observed performance variations as a result of inherent stochasticity in the training and rollout processes.

Highlights

• GSM8K Alignment Report Added: A new markdown report has been added detailing a comprehensive alignment experiment between the veRL and Trinity frameworks on the GSM8K mathematical reasoning benchmark.
• veRL Code Modifications: Modifications were made to veRL to fix a bug in DataProto.concat() for meta_info aggregation, add logic to save actor input batches, and adjust fsdp_workers.py for numerical consistency with Trinity.
• Trinity Code Modifications: Trinity was updated to load veRL-saved input batches and actor model/optimizer checkpoints, and a new VerlGSM8kWorkflow was implemented to reuse veRL's GSM8K scoring logic.
• Experimental Setup: Four distinct experiments were conducted to compare veRL and Trinity under varying conditions: exact reproduction, state-aligned training, independent training with zero regularization, and independent training with non-zero regularization.
• Experimental Results: Results confirm that veRL and Trinity exhibit nearly identical behavior under strictly controlled conditions (Experiment 1). Divergences in less constrained experiments (2-4) are attributed to expected stochasticity from vLLM sampling and independent optimization trajectories, rather than fundamental implementation differences.

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 adds a detailed benchmark report comparing veRL and Trinity on the GSM8K task. The report is well-structured and provides a comprehensive overview of the experimental setup, code modifications, and results. I've identified a few inconsistencies and areas for clarification that would improve the report's accuracy and reproducibility. My main concerns are a discrepancy in the number of experiments mentioned, inconsistent GPU configurations in the provided scripts, and some ambiguity regarding code modifications for different experimental runs. I've also noted a minor formatting suggestion.