Support GKD Liger Kernel Loss

[hjh0119]

No description provided.

[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 integrates an optional, memory-optimized loss function from the Liger Kernel into the GKDTrainer. The primary goal is to enhance training efficiency, particularly concerning memory consumption, by employing a fused Jensen-Shannon Divergence (JSD) loss that processes hidden states and output embeddings directly. This provides an alternative, more resource-conscious approach to knowledge distillation.

Highlights

• Liger Kernel Integration: Adds support for LigerFusedLinearJSDLoss from the liger_kernel library, providing an alternative, potentially more efficient, loss computation method for GKD training.
• Memory Optimization: The new Liger loss path is designed to reduce peak memory usage by directly operating on model hidden states and output heads, and explicitly releasing intermediate tensors after use.
• Conditional Loss Computation: The compute_loss method now dynamically switches between the standard JSD loss and the Liger fused JSD loss based on a configuration flag (use_liger_kernel).
• SFT Loss Constraint: An assertion has been added to ensure that SFT (Supervised Fine-Tuning) loss is not used concurrently with the Liger fused JSD loss, indicating a current incompatibility or design choice.

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 Liger Kernel Loss in the GKD trainer, which is a valuable addition for memory-efficient training. The implementation is mostly well-structured, introducing a new code path in compute_loss for the fused JSD loss. However, I've identified a critical issue with a misplaced assertion that could break existing SFT loss functionality for users not using the Liger kernel. Additionally, there's an opportunity to refactor some duplicated code to enhance maintainability. Please see my detailed comments.

[hjh0119]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces support for Liger Kernel for GKD loss calculation, which is a valuable optimization for memory efficiency. The implementation is largely well-executed. My review focuses on a potential bug in handling PEFT teacher models and a structural improvement to enhance code clarity and maintainability.