add support for qwen3 moe model EPLB

[hsliuustc]

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

This pull request add support for qwen3 moe model EPLB feature, which helps to improve the overall thoughput during LLM Serving.

#20468

Test Plan

import json
import os
import time
from vllm import LLM, SamplingParams

prompt = "Explain the theory of relativity in simple terms."

RESULT_FILE = "eplb_test_output.json"

sampling_params = SamplingParams(temperature=0.0, top_p=1.0)

def run_inference(model_path: str, enable_eplb: bool, num_redundant_experts: int = 0):
    print(f"Running inference with EPLB={enable_eplb}, redundant experts={num_redundant_experts}")
    
    llm = LLM(
        model=model_path,
        tensor_parallel_size=8,
        enable_expert_parallel=True,
        enable_eplb=enable_eplb,
        num_redundant_experts=num_redundant_experts if enable_eplb else 0,
        eplb_window_size=100,
        eplb_step_interval=100,
        enforce_eager=True,
    )
    
    outputs = llm.generate([prompt], sampling_params)
    output_text = outputs[0].outputs[0].text.strip()
    
    print(f"Output ({'EPLB' if enable_eplb else 'Normal'}):")
    print(output_text)
    
    return output_text

def save_result(key: str, value: str):
    if os.path.exists(RESULT_FILE):
        with open(RESULT_FILE, "r") as f:
            results = json.load(f)
    else:
        results = {}

    results["eplb" if key == "eplb" else "normal"] = value

    with open(RESULT_FILE, "w") as f:
        json.dump(results, f, indent=2)

    print(f"Output saved to {RESULT_FILE}")

def load_results():
    if os.path.exists(RESULT_FILE):
        with open(RESULT_FILE, "r") as f:
            return json.load(f)
    return {}

def compare_results():
    results = load_results()

    eplb_output = results["eplb"]
    normal_output = results["normal"]

    if eplb_output == normal_output:
        print("Outputs are consistent.")
    else:
        print("Outputs differ!:")
        print("\n--- EPLB Output ---\n", eplb_output)
        print("\n--- Normal Output ---\n", normal_output)

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--mode", type=str, choices=["eplb", "normal"], required=True)
    args = parser.parse_args()

    MODEL_PATH = "/workspace/models/Qwen3-235B-A22B-FP8"

    if args.mode == "eplb":
        output = run_inference(MODEL_PATH, enable_eplb=True, num_redundant_experts=32)
        save_result("eplb", output)
    else:
        output = run_inference(MODEL_PATH, enable_eplb=False)
        save_result("normal", output)

    # compare_results()

Test Result

will add the test result later

(Optional) Documentation Update

[gemini-code-assist]

Summary of Changes

Hello @hsliuustc, 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 Qwen3 Mixture of Experts (MoE) model by integrating Expert Parallel Load Balancing (EPLB) capabilities. The primary goal is to improve LLM serving throughput by enabling dynamic management of expert assignments and supporting redundant experts, thereby optimizing resource utilization and balancing computational load across devices in an expert-parallel environment.

Highlights

• EPLB Integration: This pull request introduces support for Expert Parallel Load Balancing (EPLB) within the Qwen3 Mixture of Experts (MoE) model. This is achieved by adding enable_eplb and num_redundant_experts parameters to the model's configuration and propagating them through the Qwen3MoeSparseMoeBlock and Qwen3MoeDecoderLayer down to the FusedMoE layers.
• MixtureOfExperts Protocol Implementation: The Qwen3MoeForCausalLM class now implements the MixtureOfExperts interface. This involves collecting all FusedMoE layers within the model and exposing MoE-specific properties such as num_logical_experts, num_physical_experts, and num_redundant_experts.
• Dynamic Expert Load Balancing: A new set_eplb_state method has been added to Qwen3MoeForCausalLM. This method allows for dynamic updates to the expert load view, logical-to-physical expert mapping, and logical replica counts, which are crucial for real-time load balancing across experts in an expert-parallel setup.
• Refactored Weight Loading: The load_weights mechanism for the Qwen3 MoE model has been updated to correctly handle the mapping and loading of weights for redundant experts, ensuring that the model can properly initialize with the new EPLB configuration.

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 adds support for EPLB for the Qwen3 MoE model. I found a critical bug in the type checking for MoE layers and a fragile piece of logic for identifying an example layer. My review provides suggestions to fix these issues to ensure the feature works correctly and robustly.

[gemini-code-assist]

The current logic for getting example_layer assumes the last layer of the model is always a MoE layer. This is fragile and might fail if config.num_hidden_layers is not a multiple of config.decoder_sparse_step.

        example_layer = next(layer.mlp for layer in self.model.layers
                             if isinstance(layer.mlp, Qwen3MoeSparseMoeBlock))

[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.

🚀