[Feature]Add support for SGLang in the "Model Deployment" module

[maxbetas]

1Panel Version

2.1.4

Please describe your needs or suggestions for improvements

Thank you to the 1Panel team for your excellent work on AI model deployment; the current support for Ollama and vLLM is fantastic.

As the open-source LLM ecosystem evolves, SGLang (https://github.com/sgl-project/sglang) is rapidly gaining traction as a next-generation, high-performance inference framework used by many developers and enterprises. Given 1Panel's commitment to providing comprehensive and cutting-edge productivity tools, I highly recommend adding built-in support for SGLang in the model deployment module.

Why is this needed?
While 1Panel already supports vLLM and Ollama, SGLang offers unique advantages that cover scenarios where existing solutions might not be optimal:

Exceptional Concurrency Performance (RadixAttention): SGLang's RadixAttention mechanism significantly maximizes KV Cache reuse when handling multi-turn conversations, long contexts, and complex Agent prompts. Its throughput is notably superior to vLLM in many of these scenarios.

Ultra-fast Structured Output: For AI Agent development, forced JSON structured output is a strict requirement. SGLang is deeply optimized for structured decoding, heavily outperforming other frameworks.

Native Multimodal Support: It provides out-of-the-box, high-performance support for mainstream Vision Large Language Models (VLMs).

Proposed Solution:
I suggest integrating SGLang similarly to how vLLM is currently handled—by adding an SGLang option in the engine dropdown under the Model Deployment settings.

The configuration form should ideally support the following core parameters (corresponding to SGLang's startup arguments):

Model Path (--model-path)

Service Port (--port)

Tensor Parallelism / GPU Count (--tp-size)

Context Length (--context-length)

API Key (--api-key)

On the backend, 1Panel could use SGLang's official Docker image (e.g., lmsysorg/sglang:latest) to generate the docker-compose file, spin up the service, and expose an OpenAI-compatible API.

Alternative Solutions:
Currently, users have to manually write a docker-compose.yml via 1Panel's "Container Orchestration" feature to deploy SGLang. While feasible, this method lacks the intuitive graphical management, log viewing, and convenient one-click start/stop capabilities provided by the dedicated Model Deployment module.

Additional Context:

SGLang Official Documentation: https://docs.sglang.ai/

SGLang Docker Deployment Guide: https://docs.sglang.ai/backend/docker.html

Please describe the solution you suggest

No response

Additional Information

No response

[wanghe-fit2cloud]

Thanks for the feedback. Please provide the docker-compose.yml file that you are currently using and running successfully locally. We can first review it and try to add it to the 1Panel App Store.

[maxbetas]

Thanks for the feedback. Please provide the docker-compose.yml file that you are currently using and running successfully locally. We can first review it and try to add it to the 1Panel App Store.

docker-compose.yml

[wanghe-fit2cloud]

version: "3.9"

services:
  sglang:
    image: lmsysorg/sglang:v0.5.9
    container_name: sglang
    restart: always

    ports:
      - "30000:30000"

    networks:
      - 1panel-network

    volumes:
      - ./data/modelscope:/root/.cache/modelscope
      - ./data/huggingface:/root/.cache/huggingface

    environment:
      - SGLANG_USE_MODELSCOPE=true

    entrypoint: python3 -m sglang.launch_server

    command: >
      --model-path Qwen/Qwen2.5-7B-Instruct
      --host 0.0.0.0
      --port 30000
      --tp 1
      --trust-remote-code
      --api-key XZ2ZLYJKV9rtfsjvO59f3GEtOKvZ4b5b

    ipc: host

    ulimits:
      memlock: -1
      stack: 67108864

    healthcheck:
      test: ["CMD-SHELL", "curl -f http://localhost:30000/v1/models || exit 1"]
      interval: 30s
      timeout: 10s
      retries: 5

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

networks:
  1panel-network:
    external: true

[wanghe-fit2cloud]

@maxbetas

Hi, I’m preparing an application template for SGLang in 1Panel.

1Panel applications are typically deployed on a single Linux server, so cluster communication or RDMA is usually not required.

I’ve prepared a simplified docker-compose.yml for this scenario and would appreciate it if you could help review whether this configuration looks reasonable.

Additionally, when users install SGLang from the 1Panel App Store, we would like to expose some configurable options in the installation form (similar to how vLLM is handled). For example:

• Exposed port (user customizable)
• Model path / model name
• Startup command arguments
• Data or cache directory
• API key

Do you think these parameters are appropriate to make configurable for users?

If so, would you recommend structuring the docker-compose.yml in a way that better supports user-defined variables (for example, through environment variables or command parameters)?

Thanks for your help!