Allow serving Llama4ForCausalLM directly

[sarckk]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

Addresses #18022 and #20581. Some users want to serve text-only variants of Llama4 models.

Test Plan

There is no official text-only model checkpoint at the moment, but I tested that serving still works with https://huggingface.co/trl-internal-testing/tiny-Llama4ForCausalLM

Test Result

Below command works

vllm serve trl-internal-testing/tiny-Llama4ForCausalLM --gpu-memory-utilization 0.95 --max-model-len 512 --tensor_parallel_size 1

(Optional) Documentation Update

Did not update supported_models.md yet as there is no example official HF 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 @sarckk, 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 the necessary configuration to enable vLLM to serve Llama4 text-only models. The primary change involves updating the internal model registry to properly identify and load these models, addressing a user need for this specific model variant.

Highlights

• New Model Support: I've added support for Llama4 text-only models by registering Llama4ForCausalLM in the model registry. This allows vLLM to recognize and serve these specific model variants.

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 pull request introduces support for Llama4 text-only models by adding a new model implementation file vllm/model_executor/models/llama4.py and registering it in the model registry. The implementation looks solid, covering attention mechanisms, MoE layers, and weight loading logic specific to Llama4.

My review has identified a potential correctness issue in the weight loading logic for MoE layers that could cause failures when loading certain model checkpoints. I've also noted that automated tests for this new model architecture are missing, which are important for long-term maintainability. I've provided suggestions to address both points.

[gemini-code-assist]

While adding the model to the registry is correct, this PR appears to be missing automated tests for the new Llama4ForCausalLM model. To ensure the model implementation is and remains correct, please add an entry for it in _TEXT_GENERATION_EXAMPLE_MODELS in tests/models/registry.py.

This will enable CI to run tests against this new architecture. Based on your PR description, you could use trl-internal-testing/tiny-Llama4ForCausalLM. If this model is not publicly accessible, consider creating a small, public dummy model with the same architecture for testing purposes.

[huachenheli]

Since there's no official text-only HF checkpoint, can you add a comment on what's required to enable the text-only version? e.g. Need to change architecture in config.json. Also, does one need to strip the mm part from the HF version? etc.

[DarkLight1337]

cc @22quinn @houseroad

[houseroad]

I remember @ywang96 raised some concern before?

[ywang96]

@sarckk Thanks for the PR - please see my review here #19580

[huachenheli]

@sarckk Thanks for the PR - please see my review here #19580

Can we derive text-only mode via allowed_limits? If supported modalities are limited to 0 (e.g. via --limit-mm-per-prompt), we skip loading the mm part which should have the model serve with text-only mode.

[houseroad]

Actually, using limit-mm-per-prompt seems a good idea.