Support bailing moe

[ppraneth]

SGLang vs. vLLM: Verified MMLU Benchmark on Ling-lite Model

Model	Metric	SGLang	vLLM	SGLang Advantage
inclusionAI/Ling-lite	Avg. MMLU Accuracy	70.0%	69.9%	+0.1%
	Total Latency	53.7s	94.8s	1.77x Faster

Comparison with Official MMLU Scores

For inclusionAI/Ling-lite:

• Official MMLU Score: 71.2%
• SGLang result was 1.2% lower than the official score.
• vLLM result was 1.3% lower than the official score.

---

Ling-lite Benchmark Results

Metric	SGLang	vLLM	SGLang vs. vLLM Difference
Average Accuracy	70.0%	69.9%	+0.1%
Total Latency	53.655s	94.752s	-41.097s (1.77x faster)

Detailed Subject-by-Subject Comparison (Ling-lite)

Subject	Questions	SGLang Acc	vLLM Acc	SGLang vs. vLLM Diff
Abstract Algebra	100	48.0%	49.0%	-1.0%
Anatomy	135	71.1%	71.1%	0.0%
Astronomy	152	77.0%	78.3%	-1.3%
Business Ethics	100	75.0%	74.0%	+1.0%
Clinical Knowledge	265	78.5%	77.7%	+0.8%
College Biology	144	80.6%	78.5%	+2.1%
College Chemistry	100	54.0%	54.0%	0.0%
College Computer Science	100	67.0%	64.0%	+3.0%
College Mathematics	100	45.0%	46.0%	-1.0%
College Medicine	173	74.0%	73.4%	+0.6%
College Physics	102	49.0%	48.0%	+1.0%
Computer Security	100	75.0%	76.0%	-1.0%
Conceptual Physics	235	76.2%	75.7%	+0.5%
Econometrics	114	57.9%	57.9%	0.0%
Electrical Engineering	145	69.7%	71.0%	-1.3%
Elementary Mathematics	378	70.4%	70.4%	0.0%
Formal Logic	126	55.6%	56.3%	-0.7%
Global Facts	100	43.0%	46.0%	-3.0%
High School Biology	310	89.0%	89.7%	-0.7%
High School Chemistry	203	61.6%	62.6%	-1.0%
High School Computer Science	100	84.0%	84.0%	0.0%
High School European History	165	81.2%	83.0%	-1.8%
High School Geography	198	89.4%	89.9%	-0.5%
High School Government and Politics	193	93.3%	92.7%	+0.6%
High School Macroeconomics	390	74.6%	74.9%	-0.3%
High School Mathematics	270	53.3%	52.2%	+1.1%
High School Microeconomics	238	84.9%	84.9%	0.0%
High School Physics	151	57.6%	57.6%	0.0%
High School Psychology	545	87.7%	88.3%	-0.6%
High School Statistics	216	67.1%	67.1%	0.0%
High School US History	204	80.9%	79.9%	+1.0%
High School World History	237	82.7%	82.7%	0.0%
Human Aging	223	69.1%	67.7%	+1.4%
Human Sexuality	131	79.4%	78.6%	+0.8%
International Law	121	80.2%	80.2%	0.0%
Jurisprudence	108	76.9%	76.9%	0.0%
Logical Fallacies	163	80.4%	79.8%	+0.6%
Machine Learning	112	65.2%	65.2%	0.0%
Management	103	85.4%	85.4%	0.0%
Marketing	234	86.8%	86.8%	0.0%
Medical Genetics	100	78.0%	81.0%	-3.0%
Miscellaneous	783	85.2%	84.5%	+0.7%
Moral Disputes	346	74.3%	74.6%	-0.3%
Moral Scenarios	895	36.5%	36.2%	+0.3%
Nutrition	306	76.1%	75.2%	+0.9%
Philosophy	311	75.6%	76.2%	-0.6%
Prehistory	324	79.6%	79.0%	+0.6%
Professional Accounting	282	55.0%	53.5%	+1.5%
Professional Law	1534	49.9%	50.8%	-0.9%
Professional Medicine	272	75.0%	73.9%	+1.1%
Professional Psychology	612	75.5%	74.5%	+1.0%
Public Relations	110	70.9%	69.1%	+1.8%
Security Studies	245	78.0%	78.4%	-0.4%
Sociology	201	87.6%	88.1%	-0.5%
US Foreign Policy	100	82.0%	83.0%	-1.0%
Virology	166	53.6%	54.2%	-0.6%
World Religions	171	84.2%	84.8%	-0.6%

---

Technical Configurations

SGLang Benchmark Commands

# For Ling-lite
!python -m sglang.launch_server \
  --model-path inclusionAI/Ling-lite \
  --port 30000
!python3 bench_sglang.py 

Vllm Benchmark Commands

# For Ling-lite
!python3 -m vllm.entrypoints.api_server \
  --tokenizer-mode auto \
  --model inclusionAI/Ling-lite \
  --disable-log-requests \
  --port 21000
!python3 bench_other.py  --backend vllm

Motivation

Closes #8621

Modifications

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

[gemini-code-assist]

Summary of Changes

Hello @ppraneth, 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!

I've reviewed this pull request, which primarily focuses on extending our model compatibility by adding support for Bailing Mixture-of-Experts (MoE) models. This integration includes the inclusionAI/Ling-lite and Ling-plus models, enhancing our framework's ability to handle these advanced architectures for improved performance and flexibility.

Highlights

• New Model Support: I've added comprehensive support for Bailing Mixture-of-Experts (MoE) models, specifically integrating inclusionAI/Ling-lite and Ling-plus, which significantly expands the range of advanced models our system can serve.
• Core MoE Architecture Implementation: A new Python module (python/sglang/srt/models/bailing_moe.py) has been introduced. This module defines the complete Bailing MoE model architecture, including its attention, MLP, and expert layers, adapted from the vLLM project to ensure compatibility and performance.
• Specialized Weight Loading: I've implemented custom weight loading logic within the new Bailing MoE model. This ensures correct handling of the unique parameter structures inherent to MoE models, such as stacked and expert-specific weights, facilitating seamless model integration.
• Documentation and Testing: The docs/supported_models/generative_models.md has been updated to reflect the newly supported Ling MoE models. Additionally, a dedicated test case for inclusionAI/Ling-lite has been added to test/srt/models/test_generation_models.py to validate its proper functionality and integration within our generation framework.

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 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 or fill out our survey 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 the Bailing MoE model, including the model implementation, a test case, and documentation updates. The implementation appears correct and follows the project's patterns. I've provided a few suggestions to improve the documentation's clarity and the code's maintainability. Overall, great work!

[strgrb]

@ppraneth Great Job! I'll take a look.

[strgrb]

@ant-yy Please help to take a look

[jeejeelee]

@ppraneth Can you provide your benchmark detail information? Such as GPU information, bench scripts, etc.
I'm a bit curious about the latency differences between vllm vs sglang and want to reproduce it.

[ppraneth]

@ppraneth Can you provide your benchmark detail information? Such as GPU information, bench scripts, etc. I'm a bit curious about the latency differences between vllm vs sglang and want to reproduce it.

I ran it on H100
Benchmark code:
SGLang:

import threading
import subprocess
import time
import os

def launch_server():
  
    process = subprocess.Popen(
        ["python", "-m", "sglang.launch_server", "--model-path", "inclusionAI/Ling-lite","--trust-remote-code","--port", "30000"],
        stdout=subprocess.PIPE,
        stderr=subprocess.STDOUT,
        text=True,
    )
    for line in process.stdout:
        print("[SERVER]", line.strip())

# Start server thread
server_thread = threading.Thread(target=launch_server)
server_thread.start()


!python3 bench_sglang.py 

Vllm:

import threading
import subprocess
import time

def launch_server():

    process = subprocess.Popen(
        ["python", "-m", "vllm.entrypoints.api_server","--tokenizer-mode","auto","--model", "inclusionAI/Ling-lite","--trust-remote-code","--disable-log-requests","--port", "21000"],
        stdout=subprocess.PIPE,
        stderr=subprocess.STDOUT,
        text=True
    )
    for line in process.stdout:
        print("[SERVER]", line.strip())

# Start server thread
server_thread = threading.Thread(target=launch_server)
server_thread.start()

!python3 bench_other.py --backend vllm

[jeejeelee]

Thank you very much for your reply. Can you provide bench_sglang.py and bench_other.py?

[ppraneth]

Thank you very much for your reply. Can you provide bench_sglang.py and bench_other.py?

https://github.com/sgl-project/sglang/tree/main/benchmark/mmlu

[strgrb]

LGTM!

[jeejeelee]

I am using A800 locally, then following the instructions to conduct testing. The conclusions are as follows:

• sglang

Total latency: 80.394
Average accuracy: 0.715

• vllm

Total latency: 75.417
Average accuracy: 0.718

For my environment, please refer to the attachment
test_lantency.txt

Once again, thank you for your helpful response

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!