modelscope 上PaddleOCR-VL 部署(2026年1月17日亲测可用)

PaddleOCR-VL 部署

创建时间: 2026-01-16
环境: ModelScope PAI-DSW 免费实例

环境配置

虚拟环境

wgethttp://qiniu.dywlkj.com/uv_for_linux_x86/install.sh -O -|sh

uv 安装信息

• uv 版本：0.8.22
• 已安装命令：uv和uvx

会话管理命令（推荐使用 screen 或 tmux 后台运行）

更换国内镜像源（加速下载）

方法1：临时环境变量（单次生效）

# 使用阿里云镜像源exportUV_INDEX_URL=https://mirrors.aliyun.com/pypi/simple/ uv pipinstall[包名]

常用镜像源列表

示例

wgethttp://qiniu.dywlkj.com/uv_for_linux_x86/install.sh -O -|shexportUV_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple/

or

cat<<'EOF'>uv.toml[[index]] name = "aliyun" url = "https://mirrors.aliyun.com/pypi/simple/" default = true EOFuv pipinstallmodelscope

uv 常用命令

• uv—— 快速创建虚拟环境、安装包
• uvx—— 直接运行 Python 工具（如uvx black）

虚拟环境操作

# 路径/mnt/workspace/paddleocr-vl/.venv# 创建命令mkdir-p /mnt/workspace/paddleocr-vlcd/mnt/workspace/paddleocr-vl uv venv --python3.11# 激活命令source.venv/bin/activate

PaddleOCR vs PaddleOCR-VL

核心区别

联系

PaddleOCR (基础层) ↓ └─> 文本检测/识别能力 ↓ PaddleOCR-VL (增强层) ↓ └─> 基于PaddleOCR + 多模态大模型 └─> 文档理解、问答、推理

PaddleOCR-VL 核心信息

模型规格

GitHub & 文档

• GitHub: https://github.com/PaddlePaddle/PaddleOCR-VL
• ModelScope: https://modelscope.cn/models/PaddlePaddle/PaddleOCR-VL

安装方案

推荐方案

# 激活环境source.venv/bin/activate# 安装 PaddlePaddle-GPU 3.2.0UV_HTTP_TIMEOUT=3600uv pipinstallpaddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/# 安装 PaddleOCR-VLUV_HTTP_TIMEOUT=3600uv pipinstall-U"paddleocr[doc-parser]"

备选方案（cu126 失败时）

# 使用 cu124 版本UV_HTTP_TIMEOUT=3600uv pipinstallpaddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu124/UV_HTTP_TIMEOUT=3600uv pipinstall-U"paddleocr[doc-parser]"

验证安装

# 检查 PaddlePaddle 版本和 CUDA 支持python -c"import paddle; print(paddle.__version__); print(paddle.device.cuda.device_count())"# 检查 PaddleOCR-VLpython -c"from paddleocr import PaddleOCRVL; print('PaddleOCR-VL installed successfully')"

使用 ModelScope 下载 PaddleOCR-VL 模型

frommodelscopeimportsnapshot_download MODEL_NAME="PaddlePaddle/PaddleOCR-VL"MODEL_ID="PaddleOCR-VL"SAVE_DIR=f"/mnt/workspace/models/{MODEL_ID}"if__name__=="__main__":print(f"正在从 ModelScope 下载模型:{MODEL_NAME}")print(f"保存目录:{SAVE_DIR}")print("-"*50)model_dir=snapshot_download(MODEL_NAME,cache_dir=SAVE_DIR,revision="master",)print("-"*50)print(f"模型下载完成！")print(f"模型路径:{model_dir}")

使用 ModelScope 下载 PP-DocLayoutV2 模型

frommodelscopeimportsnapshot_download MODEL_NAME="PaddlePaddle/PP-DocLayoutV2"MODEL_ID="PP-DocLayoutV2"SAVE_DIR=f"/mnt/workspace/models/{MODEL_ID}"if__name__=="__main__":print(f"正在从 ModelScope 下载模型:{MODEL_NAME}")print(f"保存目录:{SAVE_DIR}")print("-"*50)model_dir=snapshot_download(MODEL_NAME,cache_dir=SAVE_DIR,revision="master",)print("-"*50)print(f"模型下载完成！")print(f"模型路径:{model_dir}")

Key Installation Commands

# Install PaddlePaddle-GPU with CUDA 12.6 supportUV_HTTP_TIMEOUT=3600uv pipinstallpaddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/# Install PaddleOCR-VL with document parser capabilitiesUV_HTTP_TIMEOUT=3600uv pipinstall-U"paddleocr[doc-parser]"# Install vLLM server dependencies (for production deployment)paddleocr install_genai_server_deps vllm# Install flash-attention for GPU acceleration (from pre-built wheel)# Check CUDA version first: nvidia-smi | grep "CUDA Version"pipinstallhttps://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu12torch2.8cxx11abiFALSE-cp311-cp311-linux_x86_64.whl# Or compile from source (if wheel doesn't match your environment)pip uninstall -y flash_attn pipinstallpackaging ninja pipinstallflash-attn --no-build-isolation

Production Deployment Workflow

Complete Setup Flow (Verified Working)

Step 1: Environment Activation

source.venv/bin/activate

Step 2: Install Dependencies

# Core packagespaddleocr install_genai_server_deps vllm# Flash-attention (GPU acceleration)# Download wheel locally first, then installwgethttp://qiniu.dywlkj.com/deepseek-ocr-vllm/flash_attn-2.7.3%2Bcu12torch2.8cxx11abiFALSE-cp311-cp311-linux_x86_64.whl pipinstall./flash_attn-2.7.3+cu12torch2.8cxx11abiFALSE-cp311-cp311-linux_x86_64.whl# Verify installationpython -c"import flash_attn; print(flash_attn.__version__)"# Expected output: 2.7.3

Step 3: Start vLLM Server

paddlex_genai_server\--model_name PaddleOCR-VL-0.9B\--model_dir /mnt/workspace/models/PaddleOCR-VL/PaddlePaddle/PaddleOCR-VL\--backend vllm\--host0.0.0.0\--port8118

Step 4: Verify Server Status

# Check server healthcurlhttp://127.0.0.1:8118/health# List available modelscurlhttp://127.0.0.1:8118/v1/models# Expected response:# {"data":[{"id":"PaddleOCR-VL-0.9B","object":"model",...}],"object":"list"}

PaddleX Layout Parsing API

paddlex --serve\--pipeline ./config/PaddleOCR-VL.yaml\--host0.0.0.0\--port10800\--paddle_model_dir /mnt/workspace/paddleocr-vl