最强开源0.9B级OCR模型！PaddleOCR-VL本地一键部署，私密性拉满【喂饭级教程】

注意：不想看踩坑过程的朋友，可以直接全局搜："这条指令一键启动"，定位到最终的启动指令。

docker run -it --rm --gpus all -p 8118:8118 --network host ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddlex-genai-vllm-server /bin/bash

docker pull ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddlex-genai-vllm-server

python -m pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.11/flash_attn-2.8.3+cu128torch2.8-cp310-cp310-linux_x86_64.whl

paddlex_genai_server --model_name PaddleOCR-VL -0.9B  --backend vllm --port 8118 --host 0.0.0.0

paddlex_genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --port 8118 --host 0.0.0.0 --backend_config <(echo -e 'gpu-memory-utilization: 0.8')

后面我最后一次启动忘记加--backend_config <(echo -e 'gpu-memory-utilization: 0.8'，不过没报错，成功启动了，最终显存占用是6G多。

services:  paddleocr-vl-api:    image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-vl:latest-offline    container_name: paddleocr-vl-api    ports:      - 8080:8080    depends_on:      paddleocr-genai-vllm-server:        condition: service_healthy    deploy:      resources:        reservations:          devices:            - driver: nvidia              device_ids: ["0"]              capabilities: [gpu]    restart: unless-stopped    healthcheck:      test: ["CMD-SHELL", "curl -f http://localhost:8080/health || exit 1"]  paddleocr-genai-vllm-server:    image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest-offline    container_name: paddleocr-genai-vllm-server    deploy:      resources:        reservations:          devices:            - driver: nvidia              device_ids: ["0"]              capabilities: [gpu]    restart: unless-stopped    healthcheck:      test: ["CMD-SHELL", "curl -f http://localhost:8080/health || exit 1"]      start_period: 300s

docker run -d --rm --gpus all -p 8118:8118 --name paddleocr-vl-server ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest sh -c "pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.11/flash_attn-2.8.3+cu128torch2.8-cp310-cp310-linux_x86_64.whl && paddleocr genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --port 8118 --host 0.0.0.0 --backend_config <(echo -e 'gpu-memory-utilization: 0.8')"

docker run -d --rm --gpus all -p 8118:8118 --name paddleocr-vl-server ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest sh -c "paddleocr genai_server --model_name PaddleOCR-VL-0.9B --backend vllm --port 8118 --host 0.0.0.0 --backend_config <(echo -e 'gpu-memory-utilization: 0.8')"

{    "model": "PaddleOCR-VL-0.9B",    "messages": [        {            "role": "user",            "content": [                {                    "type": "text",                    "text": "请识别这张图片中的所有文本内容。"                },                {                    "type": "image_url",                    "image_url": {                        "url": "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/paddleocr_vl_demo.png"                    }                }            ]        }    ],    "max_tokens": 2048,    "temperature": 0.0}

PS：如果复制这段json用不了的朋友，可以丢给AI输出一遍，因为公众号里面复制出去可能会有格式错误，但是肉眼很难看出来。