megatron swift support KTO

[kevssim]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

megatron swift support kto.

[gemini-code-assist]

Summary of Changes

Hello @kevssim, 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 comprehensive support for Kahneman-Tversky Optimization (KTO) training within the Megatron-SWIFT framework. It includes the implementation of a dedicated KTO trainer, new command-line arguments for KTO configuration, updated documentation, and an example script to facilitate its use. The changes enable users to perform KTO-based reinforcement learning from human feedback, expanding the capabilities of the system for fine-tuning large language models.

Highlights

• New KTO Trainer Implementation: A dedicated MegatronKTOTrainer has been implemented to support Kahneman-Tversky Optimization (KTO) training, inheriting from MegatronTrainer and handling KTO-specific loss calculations and data processing.
• KTO Parameter Support: New command-line arguments (beta, desirable_weight, undesirable_weight, calculate_KL) have been added to MegatronRLHFArguments for configuring KTO training, allowing fine-grained control over the optimization process.
• Documentation Updates: Both Chinese and English documentation files (命令行参数.md and Command-line-parameters.md) have been updated to reflect the newly introduced KTO parameters and to include 'kto' as an available rlhf_type option.
• KTO Example Script: A new example shell script (kto.sh) has been added, demonstrating how to configure and run KTO training within the Megatron-SWIFT framework using the new parameters.
• Enhanced Batch Processing and RLHF Integration: Improvements were made to batch size calculation in prepare_kto_dataset, and a new get_kto_batch utility was added for KTO-specific data handling. The KTO trainer and dataset preparation are now fully integrated into the existing Megatron RLHF training pipeline.

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 support for Kahneman-Tversky Optimization (KTO) to Megatron-SWIFT. The changes include adding a new MegatronKTOTrainer, corresponding arguments, a utility for batching KTO data, an example script, and documentation updates. The implementation looks solid, but I have a few suggestions to improve code style consistency, robustness, and documentation clarity. Specifically, I've recommended renaming an argument to follow Python conventions, improving device handling in the loss function, and fixing minor issues in the documentation.