Gemma3n (Text-only) #20134

robertgshaw2-redhat · 2025-06-26T15:37:26Z

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

add new model
note: this is only the text branch. the multimodal branch will come later this week

Test Plan

vllm serve google/gemma-3n-E2B-it --disable-log-requests

MODEL=google/gemma-3n-E2B-it
PORT=8000
lm_eval --model local-completions --tasks gsm8k \
    --model_args model=$MODEL,base_url=http://127.0.0.1:$PORT/v1/completions,num_concurrent=200,max_retries=3,tokenized_requests=False \
    --limit 1000

Test Result

local-completions (model=google/gemma-3n-E2B-it,base_url=http://127.0.0.1:8000/v1/completions,num_concurrent=200,max_retries=3,tokenized_requests=False), gen_kwargs: (None), limit: 1000.0, num_fewshot: None, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.610|±  |0.0154|
|     |       |strict-match    |     5|exact_match|↑  |0.593|±  |0.0155|

(Optional) Documentation Update

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

github-actions · 2025-06-26T15:37:35Z

👋 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 @robertgshaw2-redhat, 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 the Gemma3n model, a new large language model. It involves integrating its distinct architectural components, including a novel sparse activation function and specialized layer structures, and registering it within the existing model framework.

Highlights

New Model Integration: I've added the complete architecture for the Gemma3n model, enabling its use within the system. This includes its unique layer components and overall structure.
Custom Activation Function: I've introduced GeluAndMulSparse, a specialized sparse GELU activation function that is specifically designed for and utilized by the Gemma3n model.
Advanced Layer Structures: I've implemented unique components integral to the Gemma3n architecture, such as Gemma3nAltUp for 'Alternating Updates' and Gemma3nLaurelBlock for 'Learned Augmented Residual Layers'.
Model Registry Update: I've updated the model registry to include the new Gemma3nForConditionalGeneration model, making it discoverable and loadable by the system.

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.

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

The code changes introduce a new model, Gemma3n, to the vLLM framework. The changes include adding a new activation function, GeluAndMulSparse, and the model implementation itself. There are a few suggestions to improve the code, including correcting a docstring, using f-strings for better readability, and adding a tolerance value when comparing floating-point numbers.

vllm/model_executor/models/gemma3n.py

vllm/model_executor/layers/activation.py

vllm/model_executor/models/gemma3n.py

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

ywang96

Left a few comments so PTAL - we can also add the model to documentation but indicate that only text-input is supported for now.

vllm/model_executor/layers/activation.py

vllm/model_executor/models/gemma3n.py

mgoin · 2025-06-26T18:59:18Z

It looks like models are public https://huggingface.co/collections/google/gemma-3n-685065323f5984ef315c93f4

Could we add it to the test registry? https://github.com/vllm-project/vllm/blob/04e1642e3251fc575d104c84782fafea348cfbaf/tests/models/registry.py

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

robertgshaw2-redhat · 2025-06-26T22:49:54Z

It looks like models are public https://huggingface.co/collections/google/gemma-3n-685065323f5984ef315c93f4

Could we add it to the test registry? https://github.com/vllm-project/vllm/blob/04e1642e3251fc575d104c84782fafea348cfbaf/tests/models/registry.py

I would like to get this in now so we can have it in the nightly. Ill add a test tomorrow

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

ywang96

I would like to get this in now so we can have it in the nightly. Ill add a test tomorrow

I think that's fine since this PR is a model specific PR and should be low risk (assuming the transformers version bump doesn't break anything)

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

mgoin · 2025-06-27T01:03:07Z

Basic models test is failing because there isn't a registered model. You can add one and put a skip arg in it https://buildkite.com/vllm/ci/builds/22772/steps/canvas?jid=0197aea1-2404-4b82-8219-729311822434#0197aea1-2404-4b82-8219-729311822434/214-1806

Signed-off-by: Roger Wang <hey@rogerw.me>

docs/models/supported_models.md

Signed-off-by: Roger Wang <hey@rogerw.me>

DarkLight1337 · 2025-06-27T03:41:44Z

requirements/common.txt

@@ -7,7 +7,7 @@ requests >= 2.26.0
 tqdm
 blake3
 py-cpuinfo
-transformers >= 4.51.1
+transformers >= 4.53.0


Actually wait. We should not force update transformers yet, it will break Qwen2.5-Omni. cc @Isotr0py

Is that confirmed? If so we can user users to upgrade transformers manually.

From the look of it, it seems that the blast radius is more than just on Qwen2.5-Omni. See vllm-project/vllm-ascend#1470, so I'll update this PR accordingly.

If so we can user users to upgrade transformers manually.

There is a big refactoring at config and processor in Transformers v4.53, we need to ask users to upgrade transformers manually temporarily before we add compatibility for it.

Signed-off-by: Roger Wang <hey@rogerw.me>

huytuong010101 · 2025-06-27T13:16:34Z

DO you have plan to support multimodal version as well?

ywang96 · 2025-06-27T16:14:41Z

DO you have plan to support multimodal version as well?

Yea it's planned. Stay tuned!

huytuong010101 · 2025-06-28T02:46:03Z

@ywang96 Awesome

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me>

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me> Signed-off-by: Will Eaton <weaton@redhat.com>

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me>

Larry-Gan · 2025-07-14T07:12:28Z

@ywang96 Is the multimodal version implemented yet? I don't see any PRs that may have implemented it. I was wondering if I could go data parallel with images.

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me> Signed-off-by: avigny <47987522+avigny@users.noreply.github.com>

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me>

added model

16187fe

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

gemini-code-assist bot reviewed Jun 26, 2025

View reviewed changes

robertgshaw2-redhat requested review from WoosukKwon and ywang96 June 26, 2025 15:38

gemini-code-assist bot reviewed Jun 26, 2025

View reviewed changes

vllm/model_executor/models/gemma3n.py Outdated Show resolved Hide resolved

vllm/model_executor/layers/activation.py Outdated Show resolved Hide resolved

vllm/model_executor/models/gemma3n.py Show resolved Hide resolved

address bot comments

4b02f7d

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

ywang96 reviewed Jun 26, 2025

View reviewed changes

vllm/model_executor/layers/activation.py Outdated Show resolved Hide resolved

vllm/model_executor/models/gemma3n.py Outdated Show resolved Hide resolved

vllm/model_executor/models/gemma3n.py Outdated Show resolved Hide resolved

robertgshaw2-redhat changed the title ~~added model~~ Gemma3n Jun 26, 2025

update version

ed2e3f0

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

mergify bot added the ci/build label Jun 26, 2025

address comment

7bc8f13

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

ywang96 approved these changes Jun 26, 2025

View reviewed changes

ywang96 added the ready ONLY add when PR is ready to merge/full CI is needed label Jun 26, 2025

ywang96 changed the title ~~Gemma3n~~ Gemma3n (Text-only) Jun 26, 2025

updated

c3b3f16

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com>

robertgshaw2-redhat requested a review from hmellor as a code owner June 26, 2025 23:43

mergify bot added the documentation Improvements or additions to documentation label Jun 26, 2025

robertgshaw2-redhat enabled auto-merge (squash) June 26, 2025 23:44

add to registry

1bf7eee

Signed-off-by: Roger Wang <hey@rogerw.me>

ywang96 requested a review from DarkLight1337 as a code owner June 27, 2025 01:03

mgoin approved these changes Jun 27, 2025

View reviewed changes

ywang96 removed the request for review from DarkLight1337 June 27, 2025 01:04

text-only

aea70dd

Signed-off-by: Roger Wang <hey@rogerw.me>

jeejeelee approved these changes Jun 27, 2025

View reviewed changes

DarkLight1337 requested changes Jun 27, 2025

View reviewed changes

docs/models/supported_models.md Outdated Show resolved Hide resolved

add to registry

f914b01

Signed-off-by: Roger Wang <hey@rogerw.me>

DarkLight1337 approved these changes Jun 27, 2025

View reviewed changes

Roger Wang added 4 commits June 26, 2025 20:38

update doc

55360b0

Signed-off-by: Roger Wang <hey@rogerw.me>

remove

add8f13

Signed-off-by: Roger Wang <hey@rogerw.me>

remove pp

d31324a

Signed-off-by: Roger Wang <hey@rogerw.me>

typo

5ec4473

Signed-off-by: Roger Wang <hey@rogerw.me>

DarkLight1337 reviewed Jun 27, 2025

View reviewed changes

Roger Wang added 2 commits June 26, 2025 20:49

update

8df9d1a

Signed-off-by: Roger Wang <hey@rogerw.me>

add note

fe11f36

Signed-off-by: Roger Wang <hey@rogerw.me>

Isotr0py approved these changes Jun 27, 2025

View reviewed changes

ywang96 disabled auto-merge June 27, 2025 05:51

ywang96 enabled auto-merge (squash) June 27, 2025 05:51

ywang96 merged commit d1c956d into vllm-project:main Jun 27, 2025
74 checks passed

DarkLight1337 mentioned this pull request Jun 27, 2025

[New Model]: Gemma 3n support #18476

Closed

1 task

xjpang pushed a commit to xjpang/vllm that referenced this pull request Jun 30, 2025

Gemma3n (Text-only) (vllm-project#20134)

ec6f3d8

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me>

wseaton pushed a commit to wseaton/vllm that referenced this pull request Jun 30, 2025

Gemma3n (Text-only) (vllm-project#20134)

58f4258

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me>

juburr mentioned this pull request Jul 21, 2025

[Bug]: When using gemma-3n in Apple Silicon I get a NotImplementedError #20521

Open

1 task

googlercolin pushed a commit to googlercolin/vllm that referenced this pull request Aug 29, 2025

Gemma3n (Text-only) (vllm-project#20134)

c42783c

Signed-off-by: rshaw@neuralmagic.com <robertgshaw2@gmail.com> Signed-off-by: Roger Wang <hey@rogerw.me> Co-authored-by: Roger Wang <hey@rogerw.me>

Uh oh!

Gemma3n (Text-only) #20134

Gemma3n (Text-only) #20134

Uh oh!

Conversation

robertgshaw2-redhat commented Jun 26, 2025 • edited by github-actions bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Essential Elements of an Effective PR Description Checklist

Purpose

Test Plan

Test Result

(Optional) Documentation Update

Uh oh!

github-actions bot commented Jun 26, 2025

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ywang96 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mgoin commented Jun 26, 2025

Uh oh!

robertgshaw2-redhat commented Jun 26, 2025

Uh oh!

ywang96 left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mgoin commented Jun 27, 2025

Uh oh!

Uh oh!

DarkLight1337 Jun 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ywang96 Jun 27, 2025

Choose a reason for hiding this comment

Uh oh!

DarkLight1337 Jun 27, 2025

Choose a reason for hiding this comment

Uh oh!

ywang96 Jun 27, 2025

Choose a reason for hiding this comment

Uh oh!

Isotr0py Jun 27, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

huytuong010101 commented Jun 27, 2025

Uh oh!

ywang96 commented Jun 27, 2025

Uh oh!

huytuong010101 commented Jun 28, 2025

Uh oh!

Larry-Gan commented Jul 14, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

robertgshaw2-redhat commented Jun 26, 2025 •

edited by github-actions bot

Loading

ywang96 left a comment •

edited

Loading

ywang96 left a comment •

edited

Loading

DarkLight1337 Jun 27, 2025 •

edited

Loading

Larry-Gan commented Jul 14, 2025 •

edited

Loading