[Feature] Add support for MoE models in the calibration-free RTN-based quantization

[sakogan]

This PR enhances the work started in #18768.

It adds changes required to support MoE models (e.g., Llama4) quantized with RTN.
It also expands the relevant unit test to include an MoE model.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @sakogan, 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 Mixture-of-Experts (MoE) models within the existing calibration-free Round-to-Nearest (RTN) based quantization framework. My changes enable efficient quantization of MoE layers by adapting the core quantization and dequantization logic to handle multi-dimensional expert weights and integrating a new quantization method specifically for MoE architectures.

Highlights

• MoE Model Quantization: I've added a new RTNMoEMethod to enable calibration-free RTN-based quantization specifically for FusedMoE layers. This allows MoE models, such as Jamba-tiny-dev, to leverage this efficient quantization scheme.
• Quantization Logic Adaptation: The core rtn_quantize and rtn_dequantize functions have been updated to correctly handle 3-dimensional tensors, which are essential for processing batched expert weights in MoE models. Additionally, the quantized data representation has been transitioned from signed int8 to unsigned uint8 for improved consistency and range utilization.
• Weight Handling Refinement: I've enhanced the RTNTensor by adding __getitem__ support and adjusting its shape property to correctly interpret 3D tensor dimensions. A new fix_weights helper function was also introduced to streamline the process of converting RTNParameter instances to standard Parameter objects after model loading, improving code reusability and clarity.
• Expanded Test Coverage: To ensure the robustness of the new MoE quantization support, I've included an MoE model (ai21labs/Jamba-tiny-dev) in the existing unit tests. This verifies that the quantization method functions as expected across different model architectures.

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 is currently in preview and 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 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 PR adds support for Mixture-of-Experts (MoE) models to the RTN quantization method. I've identified a critical bug in a tensor reshape operation that will cause a runtime error.

[gemini-code-assist]

The expression old_weight.shape[1] * 2 results in a dimension that is four times data.shape[1], which will cause a RuntimeError due to a mismatch in the number of elements during the reshape operation.

The * 2 seems to be a typo. Removing it will make the reshape operation mathematically valid (assuming the last dimension of the weight tensor is even). Please verify the correct shape expected by the underlying kernel.

Suggested change

	data = data.reshape(old_weight.shape[0], old_weight.shape[1] * 2, -1)
	data = data.reshape(old_weight.shape[0], old_weight.shape[1], -1)