[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 significantly enhances the calibration-free RTN-based quantization framework by adding comprehensive support for Mixture-of-Experts (MoE) models. It introduces dedicated methods and refines existing quantization logic to seamlessly integrate MoE layers, ensuring that these complex models can leverage the benefits of RTN quantization.

Highlights

• MoE Model Support: Introduced a new RTNMoEMethod class to specifically handle the quantization and dequantization of weights within FusedMoE layers, enabling RTN quantization for Mixture-of-Experts architectures. The RTNConfig now correctly identifies and applies this method to MoE layers.
• Quantization Logic Refinement: Updated the core rtn_quantize and rtn_dequantize functions to correctly process 3-dimensional tensors (e.g., [num_experts, hidden_size, intermediate_size]), which are common in MoE layers. This includes adjustments to group size calculations, clamping ranges, and the internal representation of quantized weights, transitioning to unsigned 8-bit integers (torch.uint8).
• Weight Handling Improvements: Refactored the logic for converting RTNParameter instances to standard torch.nn.Parameter after loading into a new, reusable fix_weights helper function. This improves code maintainability and applies to both standard linear layers and new MoE layers.
• Expanded Test Coverage: Added an MoE model (ai21labs/Jamba-tiny-dev) to the RTN quantization unit tests, ensuring that the new functionality for MoE models is properly validated.

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 introduces RTN quantization support for MoE models. I've identified a critical issue in the fix_weights function related to tensor reshaping, and two high severity issues in rtn_quantize and rtn_dequantize related to clamping and centering of quantized values. Addressing these issues is crucial for the correctness and stability of the quantization implementation.

[gemini-code-assist]

The dequantization process subtracts q_range // 2 from the data data -= q_range // 2 only when num_bits == 8. This subtraction should also be applied when num_bits == 4 to correctly center the dequantized values around zero.

data.copy_(tensor)
        data -= q_range // 2

[sakogan]

That's incorrect. The shift (subtraction) happens also when num_bits == 4: https://github.com/sakogan/vllm/blob/78dcad8246f710c2ef9fe712f67894c02cb7d4d3/vllm/model_executor/layers/quantization/rtn.py#L418-L419

[mgoin]

Interesting, these changes look reasonable to me. Nice work integrating and looking forward to future work.