Qwen FP8 ModelOPT support #20734

jingyu-ml · 2025-07-10T05:14:46Z

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 the Qwen FP8 support from ModelOPT

Test Plan

TBD

Test Result

TBD

(Optional) Documentation Update

github-actions · 2025-07-10T05:14:54Z

👋 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 @jingyu-ml, 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 expands vLLM's quantization capabilities by integrating full support for ModelOPT FP8 quantized models, with a particular focus on Qwen models. It introduces new quantization methods for linear and Mixture-of-Experts (MoE) layers, refines the weight loading process to accommodate ModelOPT's specific checkpoint formats and scale handling, and includes model-specific adaptations to ensure seamless loading and execution of these quantized models. The changes aim to broaden the range of supported models and improve loading efficiency for quantized architectures.

Highlights

ModelOPT FP8 Quantization Support: Added comprehensive support for ModelOPT FP8 quantization, enabling vLLM to load and run models quantized with this method. This includes identifying ModelOPT FP8 from Hugging Face configurations and handling its specific data formats.
Mixture-of-Experts (MoE) Layer Quantization: Introduced ModelOptFp8MoEMethod to specifically handle FP8 quantization for MoE layers. This new method manages the creation of FP8 weights and scales, processes scales after loading (e.g., combining w1 and w3 scales), and applies the quantization during the forward pass.
Flexible Weight Loading Enhancements: Improved the weight loading mechanism across various models to dynamically adapt to different ModelOPT weight scale patterns (e.g., weight_scale vs. weight_scale_2) and to correctly handle complex expert scale broadcasting from single values to multiple experts. This includes updates to llama4, mllama4, qwen2, and qwen3_moe.
Performance Optimizations for MoE: Introduced supports_moe_loading flags to relevant weight loaders (FusedMoE, experts_int8, moe_wna16) to explicitly mark them as MoE-aware. This optimization helps avoid expensive runtime reflection during model loading, improving efficiency for MoE architectures.
KV Cache Scale Remapping: Enhanced the maybe_remap_kv_scale_name utility to correctly remap KV cache scale names, including support for qkv_proj parameters, ensuring accurate loading of attention-related scales in quantized models.

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

This pull request adds support for Qwen models with FP8 quantization using ModelOPT. The changes span several files, including configuration, model layers, and weight loading utilities. Key additions include logic to parse ModelOPT-specific quantization configurations and a new ModelOptFp8MoEMethod for handling MoE layers. Overall, the changes are well-structured to support the new quantization format.

gemini-code-assist · 2025-07-10T05:17:29Z

vllm/config.py

+            if quant_cfg["producer"]["name"].lower() == "modelopt":
+                if "quant_algo" in quant_cfg.keys() and quant_cfg["quant_algo"].lower() == "fp8":
+                    quant_cfg = {"quant_method": "modelopt"}


The direct access to quant_cfg["producer"]["name"] is unsafe and could raise a KeyError if the keys are not present in the quant_cfg dictionary. It's better to use .get() for safer access. Additionally, the nested if statements can be combined for improved readability.

Suggested change

if quant_cfg["producer"]["name"].lower() == "modelopt":

if "quant_algo" in quant_cfg.keys() and quant_cfg["quant_algo"].lower() == "fp8":

quant_cfg = {"quant_method": "modelopt"}

if (

quant_cfg is not None

and quant_cfg.get("producer", {}).get("name", "").lower() == "modelopt"

and quant_cfg.get("quant_algo", "").lower() == "fp8"

):

quant_cfg = {"quant_method": "modelopt"}

gemini-code-assist · 2025-07-10T05:17:29Z

vllm/model_executor/layers/quantization/modelopt.py

+            quant_method = cls.get_from_keys(config, ["quant_algo"])
+            kv_cache_quant_method = cls.get_from_keys(config, ["kv_cache_scheme"])
+            exclude_modules = cls.get_from_keys(config, ["ignore"])
+        except:


Using a bare except: is risky because it catches all exceptions, including system-exiting ones like SystemExit and KeyboardInterrupt. This can hide bugs and make the program difficult to terminate. It's better to catch a more specific exception, or at least Exception.

Suggested change

except:

except Exception:

gemini-code-assist · 2025-07-10T05:17:29Z

vllm/model_executor/layers/fused_moe/layer.py

+
+# Mark the FusedMoE weight_loader as supporting MoE-specific parameters
+# to avoid expensive runtime reflection in model loading code
+FusedMoE.weight_loader.supports_moe_loading = True  # type: ignore[attr-defined]


Consider using torch.inference_mode() or torch.no_grad() to disable gradient calculation during inference, potentially reducing memory consumption and improving performance.

Suggested change

FusedMoE.weight_loader.supports_moe_loading = True # type: ignore[attr-defined]

with torch.inference_mode():

FusedMoE.weight_loader.supports_moe_loading = True # type: ignore[attr-defined]

mergify · 2025-07-12T06:11:58Z

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @jingyu-ml.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

Edwardf0t1 and others added 27 commits July 9, 2025 06:14

resolve conflict

6927b02

bugfix

b45972e

handle language_model. prefix

bf96528

fix issue in fused_experts calling

cb20cd1

minor

1f61802

update ModelOptFp8Config, handle prefix in mllama4 weight loading, debug

0fb23e1

debug, handle kv scales

03d2b3b

fix kv scale name matching issue

93d7185

update, debug

d154fe1

cleanup

782c018

fix format

b78b191

resolve conflict

22745f3

debug

1c5acec

handle eplb in ModelOptFp8MoEMethod

b10782d

broadcasting BMM experts scales

8265287

cleanup

59190ea

some refactor and cleanup

7a6fc84

refactor Llama4ForConditionalGeneration.load_weights

47a47a9

resolve conflict

eec1daf

format and linter error fix

770bc24

simplify ModelOptFp8MoEMethod to avoid mypy error

770890a

resolve conflict

1beecbf

format fix

0b98a7f

fix mypy error

cc44385

Merge remote-tracking branch 'vllm/main' into zhiyu/llama4-fp8-modelopt

1ece491

Merge remote-tracking branch 'vllm/main' into jingyux/dev-qwen-fp8

767358e

add qwen fp8 modelopt support

1206f33

jingyu-ml requested review from simon-mo, WoosukKwon and youkaichao as code owners July 10, 2025 05:14

jingyu-ml requested review from robertgshaw2-redhat, mgoin, tlrmchlsmth, houseroad and hmellor as code owners July 10, 2025 05:14

jingyu-ml marked this pull request as draft July 10, 2025 05:14

mergify bot added llama Related to Llama models qwen Related to Qwen models labels Jul 10, 2025

gemini-code-assist bot reviewed Jul 10, 2025

View reviewed changes

mergify bot added the needs-rebase label Jul 12, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

Qwen FP8 ModelOPT support #20734

Qwen FP8 ModelOPT support #20734

jingyu-ml commented Jul 10, 2025 •

edited by github-actions bot

Loading

Uh oh!

github-actions bot commented Jul 10, 2025

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jul 10, 2025

Uh oh!

gemini-code-assist bot Jul 10, 2025

Uh oh!

gemini-code-assist bot Jul 10, 2025

Uh oh!

mergify bot commented Jul 12, 2025

Uh oh!

Uh oh!

	FusedMoE.weight_loader.supports_moe_loading = True # type: ignore[attr-defined]
	with torch.inference_mode():
	FusedMoE.weight_loader.supports_moe_loading = True # type: ignore[attr-defined]

Uh oh!

Qwen FP8 ModelOPT support #20734

Are you sure you want to change the base?

Qwen FP8 ModelOPT support #20734

Conversation

jingyu-ml commented Jul 10, 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 Jul 10, 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!

gemini-code-assist bot Jul 10, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jul 10, 2025

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jul 10, 2025

Choose a reason for hiding this comment

Uh oh!

mergify bot commented Jul 12, 2025

Uh oh!

Uh oh!

jingyu-ml commented Jul 10, 2025 •

edited by github-actions bot

Loading