Skip to content

Support Llama 4 for cutlass_moe_fp4 #20453

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 5 commits into from
Jul 9, 2025
Merged

Support Llama 4 for cutlass_moe_fp4 #20453

merged 5 commits into from
Jul 9, 2025

Conversation

mgoin
Copy link
Member

@mgoin mgoin commented Jul 3, 2025 •
edited by github-actions bot
Loading

Purpose

Implements apply_router_weight_on_input for the cutlass_moe_fp4 implementation in order to support Llama 4 NVFP4

Test Plan

Model evaluation

Test Result

vllm (pretrained=Llama-4-Scout-17B-16E-NVFP4,max_model_len=4096,tensor_parallel_size=1,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8120|±  |0.0108|
|     |       |strict-match    |     5|exact_match|↑  |0.8105|±  |0.0108|

vllm (pretrained=Llama-4-Scout-17B-16E-NVFP4,max_model_len=4096,tensor_parallel_size=1,enforce_eager=True,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: auto
|  Tasks   |Version|Filter|n-shot|Metric|   |Value |   |Stderr|
|----------|------:|------|-----:|------|---|-----:|---|-----:|
|winogrande|      1|none  |     5|acc   |↑  |0.7782|±  |0.0117|

@mgoin
Support Llama 4 for cutlass_moe_fp4
0710b7d 
Signed-off-by: mgoin <mgoin64@gmail.com>
@github-actions GitHub Actions
Copy link

github-actions bot commented Jul 3, 2025

👋 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[bot]
gemini-code-assist bot reviewed Jul 3, 2025
View reviewed changes
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Summary of Changes

Hello @mgoin, 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 a significant enhancement to the cutlass_moe_fp4 implementation by adding support for Llama 4 NVFP4 models. This is achieved through a new parameter that controls how router weights are applied, allowing them to be applied directly to the input activations, which is a specific requirement for these models. The changes also include necessary compatibility checks and integrations within the existing MoE execution framework.

Highlights

  • Llama 4 NVFP4 Support: The core cutlass_moe_fp4 function now includes a new boolean parameter, apply_router_weight_on_input, which is essential for supporting Llama 4 models that utilize NVFP4 quantization.
  • Conditional Weight Application Logic: The implementation of cutlass_moe_fp4 has been updated to conditionally apply topk_weights. If apply_router_weight_on_input is true (and topk=1), the weights are applied directly to the input activations; otherwise, they are applied to the output, ensuring correct behavior and preventing redundant scaling.
  • Marlin Backend Compatibility: Assertions have been added to explicitly disallow the use of the apply_router_weight_on_input feature when the Marlin backend is enabled for MoE operations, indicating a current incompatibility or limitation.
  • Integration with MoE Methods: The new apply_router_weight_on_input parameter is now correctly propagated and passed to the cutlass_moe_fp4 function from the CompressedTensorsW4A4MoeMethod and ModelOptNvFp4FusedMoE classes, ensuring the feature is accessible through the higher-level MoE execution paths.
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. ↩

@mergify mergify bot added the llama Related to Llama models label Jul 3, 2025
gemini-code-assist[bot]
gemini-code-assist bot reviewed Jul 3, 2025
View reviewed changes
Copy link
Contributor

@gemini-code-assist gemini-code-assist bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request introduces support for apply_router_weight_on_input for the cutlass_moe_fp4 implementation to support Llama 4 NVFP4. To improve the robustness of the code, I suggest replacing assert statements with explicit exception handling (NotImplementedError). This ensures that the checks are always performed, even in production environments where assertions might be disabled.

@mgoin
Use proper out_dtype
9fd5a86 
Signed-off-by: mgoin <mgoin64@gmail.com>
@mgoin mgoin added the ready ONLY add when PR is ready to merge/full CI is needed label Jul 8, 2025
LucasWilkinson
LucasWilkinson approved these changes Jul 9, 2025
View reviewed changes
Copy link
Collaborator

@LucasWilkinson LucasWilkinson left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@LucasWilkinson LucasWilkinson merged commit 31b96d1 into vllm-project:main Jul 9, 2025
72 checks passed
@mgoin mgoin deleted the cutlass_moe_fp4-apply_router_weight_on_input branch July 9, 2025 20:00
Chen-zexi pushed a commit to Chen-zexi/vllm that referenced this pull request Jul 13, 2025
@mgoin @Chen-zexi
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
c08676a 
Signed-off-by: mgoin <mgoin64@gmail.com>
patrickvonplaten pushed a commit to patrickvonplaten/vllm that referenced this pull request Jul 15, 2025
@mgoin @patrickvonplaten
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
6716083 
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Patrick von Platen <patrick.v.platen@gmail.com>
LyrisZhong pushed a commit to LyrisZhong/vllm that referenced this pull request Jul 23, 2025
@mgoin @LyrisZhong
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
bcf18b7 
Signed-off-by: mgoin <mgoin64@gmail.com>
avigny pushed a commit to avigny/vllm that referenced this pull request Jul 31, 2025
@mgoin @avigny
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
905250c 
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: avigny <47987522+avigny@users.noreply.github.com>
Pradyun92 pushed a commit to Pradyun92/vllm that referenced this pull request Aug 6, 2025
@mgoin
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
2c25834 
Signed-off-by: mgoin <mgoin64@gmail.com>
npanpaliya pushed a commit to odh-on-pz/vllm-upstream that referenced this pull request Aug 6, 2025
@mgoin @npanpaliya
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
db1b39d 
Signed-off-by: mgoin <mgoin64@gmail.com>
jinzhen-lin pushed a commit to jinzhen-lin/vllm that referenced this pull request Aug 9, 2025
@mgoin @jinzhen-lin
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
4e48a9a 
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>
taneem-ibrahim pushed a commit to taneem-ibrahim/vllm that referenced this pull request Aug 14, 2025
@mgoin @taneem-ibrahim
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
ac880fa 
Signed-off-by: mgoin <mgoin64@gmail.com>
diegocastanibm pushed a commit to diegocastanibm/vllm that referenced this pull request Aug 15, 2025
@mgoin @diegocastanibm
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
b5b845e 
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Diego-Castan <diego.castan@ibm.com>
epwalsh pushed a commit to epwalsh/vllm that referenced this pull request Aug 27, 2025
@mgoin @epwalsh
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
4be9876 
Signed-off-by: mgoin <mgoin64@gmail.com>
googlercolin pushed a commit to googlercolin/vllm that referenced this pull request Aug 29, 2025
@mgoin @googlercolin
Support Llama 4 for cutlass_moe_fp4 (vllm-project#20453)
ee8aa61 
Signed-off-by: mgoin <mgoin64@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@LucasWilkinson LucasWilkinson LucasWilkinson approved these changes

@robertgshaw2-redhat robertgshaw2-redhat Awaiting requested review from robertgshaw2-redhat robertgshaw2-redhat is a code owner

@tlrmchlsmth tlrmchlsmth Awaiting requested review from tlrmchlsmth tlrmchlsmth is a code owner

+1 more reviewer

@gemini-code-assist gemini-code-assist[bot] gemini-code-assist[bot] left review comments

Reviewers whose approvals may not affect merge requirements
Assignees
No one assigned
Labels
llama Related to Llama models ready ONLY add when PR is ready to merge/full CI is needed
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@mgoin @LucasWilkinson