Deployment Guide

This chapter describes the deployment process for the Multimodal Large Model. Confirm that the server meets the software and hardware requirements, then download the scripts, model weights, and image files required for deployment. Run the deployment script and verify that the service is running properly.

System Requirements

  • GPU: An NVIDIA GPU with CUDA support. At least 80 GB of VRAM is recommended.
  • Disk: At least 200 GB of available space. The model files are approximately 60 GB, and installation and extraction require additional space.
  • Software: Docker 20.10 or later, Docker Compose 2.0 or later, and nvidia-container-toolkit.

Deployment Restrictions: To ensure the stability of the DJI FlightHub 2 cluster, the Multimodal Large Model must be deployed on a separate server outside the cluster. It must not be deployed on a DJI FlightHub 2 cluster node.

File Download

FileDescriptionDownload Link
deploy-v1.0.0.zipDeployment script (required)Download
model-v1.0.0.tar.gzModel weights (required)Download
docker-image-v1.0.0.tar.gzDocker image (required)Download
checksums.md5Checksum file (optional)Download

Quick Deployment

Step 1: Download and Organize Files

After downloading the required files, store them in the following structure:

qwen3-vl-deployment/            任意目录名
├── deploy-v1.0.0.zip           部署脚本包
├── model-v1.0.0.tar.gz         模型文件
└── docker-image-v1.0.0.tar.gz  Docker 镜像

Step 2: Unzip the Deployment Package

cd qwen3-vl-deployment
unzip deploy-v1.0.0.zip

Directory structure after extraction:

qwen3-vl-deployment/
├── docker-compose.yml
├── init.sh
├── .env
├── README.md
├── model-v1.0.0.tar.gz
└── docker-image-v1.0.0.tar.gz

Step 3: Perform Deployment

chmod +x init.sh
./init.sh --offline

Step 4: Verify the Service

  1. Check service status.

    curl http://localhost:8000/health

  2. Test the chat interface.

curl -X POST http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-VL-32B-Instruct",
    "messages": [{"role": "user", "content": "你好"}]
  }'

Note: Initial startup takes 2 to 5 minutes to load the model. View progress using docker-compose logs -f.