docker/docker-compose.yml

version: '3.8'

services:
  dots-ocr-server:
    image: dots-ocr:latest
    container_name: dots-ocr-container
    ports:
      - "8000:8000"
    volumes:
      #download model to local，model url：https://www.modelscope.cn/models/rednote-hilab/dots.ocr
      - ./model/dots.ocr:/workspace/weights/DotsOCR
    environment:
      - PYTHONPATH=/workspace/weights:$PYTHONPATH
    deploy:
      resources:
        reservations:
          devices:
            - capabilities: [gpu]
              device_ids: ['0']
    entrypoint: /bin/bash
    command:
      - -c
      - |
        set -ex;
        echo '--- Starting setup and server ---';
        echo 'Modifying vllm entrypoint...';
        # This sed command patches the vllm entrypoint script to import the custom modeling code.
        sed -i '/^from vllm\.entrypoints\.cli\.main import main/a from DotsOCR import modeling_dots_ocr_vllm' $(which vllm) && \
        echo 'vllm script after patch:';
        # Show the patched part of the vllm script for verification.
        grep -A 1 'from vllm.entrypoints.cli.main import main' $(which vllm) && \
        echo 'Starting server...';
        # Use 'exec' to replace the current shell process with the vllm server,
        # ensuring logs are properly forwarded to Docker's standard output.
        exec vllm serve /workspace/weights/DotsOCR \
            --tensor-parallel-size 1 \
            --gpu-memory-utilization 0.8 \
            --chat-template-content-format string \
            --served-model-name dotsocr-model \
            --trust-remote-code