[feat] Support EAGLE3 for Qwen2

[KerwinKai]

Motivation

Follow the pr #7745 and sgl-project/SpecForge#68, i hope to add support eagle3 for qwen2 and qwq model.

[gemini-code-assist]

Summary of Changes

Hello @KerwinKai, 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 core support for the EAGLE3 feature within the Qwen2 and Qwen2-MoE large language models. It modifies the model's inference path to capture and process auxiliary hidden states, which is a prerequisite for advanced decoding techniques, and provides a new API for configuring this behavior.

Highlights

• EAGLE3 Support for Qwen2 Models: I've implemented support for the EAGLE3 feature by enabling the capture and utilization of auxiliary hidden states within the forward pass of both the Qwen2 and Qwen2-MoE models. This is crucial for functionalities like speculative decoding.
• Configurable Layer Capture: A new method, set_eagle3_layers_to_capture, has been added to Qwen2ForCausalLM and Qwen2MoeForCausalLM. This allows for flexible configuration of which intermediate layers' hidden states are captured for EAGLE3, providing sensible default layer selections if none are specified.

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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

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 issue 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 or fill out our survey to provide feedback.

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 introduces support for EAGLE3 speculative decoding in Qwen2 and Qwen2-MoE models. The changes involve adding a method to specify which hidden layers to capture and processing these auxiliary hidden states during the forward pass. The review focuses on improving code robustness and maintainability, highlighting potential AttributeError exceptions, off-by-one errors in layer indexing, and opportunities to reduce code duplication.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!