Support for multi-layer Eagle3 in SGLang

[xiaoxi-s]

Motivation

We want to support more than one layer of decoders in draft models of Eagle3. This PR implements all decoder layers to accept both input_embeds and hidden_states as input. The goal is to compare with #18175 according to SpecForge benchmarks and merge only the better PR.

Modifications

• The exception of EAGLE3 only supports 1 layer of decoder is removed.
• Multiple decoder layers can be initialized in the Eagle3 draft model with the config.num_hidden_layers key.
• Some proper renames for the Eagle3 specific classes.

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @xiaoxi-s, 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 significantly enhances the SGLang framework by introducing comprehensive support for multi-layer Eagle3 models. The changes involve a refactoring of the LlamaDecoderLayer and LlamaModel components to correctly handle and iterate through an arbitrary number of hidden layers, thereby lifting the previous architectural limitation to a single layer. This allows for greater model complexity and flexibility in Eagle3 implementations.

Highlights

• Multi-layer Support: The primary change enables the Eagle3 model to support multiple hidden layers by refactoring the LlamaModel to use an nn.ModuleList for its decoder layers instead of a single layer.
• LlamaDecoderLayer Refactoring: The forward method of LlamaDecoderLayer was updated to remove the residual parameter from its signature and explicitly apply the residual connection after the MLP operation.
• Constraint Removal: The ValueError check in Eagle3Model.__init__ that previously restricted Eagle3 models to only one hidden layer has been removed, allowing for flexible layer configurations.

🧠 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 adds support for multi-layer Eagle3 models in SGLang. The changes primarily involve updating LlamaModel to handle multiple decoder layers instead of a single one. This is achieved by replacing self.midlayer with self.midlayers (a ModuleList) and looping through them in the forward pass. The constraint limiting the model to a single layer has also been correctly removed.

My review identified a critical bug in the forward pass where a residual connection is applied twice, which would lead to incorrect model outputs. I have provided a detailed comment and a code suggestion to rectify this issue. The other changes appear correct and well-aligned with the goal of this pull request.

[dongyibo]

Hello, if trained on the same dataset, how much accuracy improvement can be expected when comparing a multi-layer Eagle model to a single-layer Eagle model? Are there any comparative data on inference performance? Thank you

[xiaoxi-s]

Hello, if trained on the same dataset, how much accuracy improvement can be expected when comparing a multi-layer Eagle model to a single-layer Eagle model? Are there any comparative data on inference performance? Thank you

There are indeed some improvements on accept_length by training on the Shared GPT dataset with multiple decoder layers. Please see the "Additional benchmark results" at the end of this comment from SpecForge, as this project is SpecForge's motivation to support more diverse draft models under the context of EAGLE3.