centos7使用gpu加速的MinerU

https://mineru.readthedocs.io/zh-cn/latest/user_guide/install/boost_with_cuda.html
由于官方只有ubantu的安装教程，并没有基于centos7的，故需要自己修改命令安装并使用。

在运行此 Docker 容器之前，您可以使用以下命令检查您的设备是否支持 Docker 上的 CUDA 加速。

docker run --rm --gpus=all nvidia/cuda:12.1.0-base-centos7 nvidia-smi

注意cuda的版本需要和nvidia-smi中显示的一致
验证结果：

那就不用docker，直接新建环境并在conda环境中使用gpu加速即可。

1.安装 magic-pdf

conda create -n mineru python=3.10
conda activate mineru
pip install -U "magic-pdf[full]" --extra-index-url https://wheels.myhloli.com

2.下载模型

将download_models_hf.py修改为使用modelscope下载

import json
import os

import requests
from modelscope import snapshot_download


def download_json(url):
    # 下载JSON文件
    response = requests.get(url)
    response.raise_for_status()  # 检查请求是否成功
    return response.json()


def download_and_modify_json(url, local_filename, modifications):
    if os.path.exists(local_filename):
        data = json.load(open(local_filename))
        config_version = data.get('config_version', '0.0.0')
        if config_version < '1.1.1':
            data = download_json(url)
    else:
        data = download_json(url)

    # 修改内容
    for key, value in modifications.items():
        data[key] = value

    # 保存修改后的内容
    with open(local_filename, 'w', encoding='utf-8') as f:
        json.dump(data, f, ensure_ascii=False, indent=4)


if __name__ == '__main__':
    # ModelScope 模型路径
    mineru_patterns = [
        "models/Layout/LayoutLMv3/*",
        "models/Layout/YOLO/*",
        "models/MFD/YOLO/*",
        "models/MFR/unimernet_small_2501/*",
        "models/TabRec/TableMaster/*",
        "models/TabRec/StructEqTable/*",
    ]
    model_dir = snapshot_download('opendatalab/PDF-Extract-Kit-1.0', allow_patterns=mineru_patterns)

    layoutreader_pattern = [
        "*.json",
        "*.safetensors",
    ]
    #layoutreader_model_dir = snapshot_download('hantian/layoutreader', allow_patterns=layoutreader_pattern)
    layoutreader_model_dir = snapshot_download('zxyayase/layoutreader', allow_patterns=layoutreader_pattern)


    model_dir = model_dir + '/models'
    print(f'model_dir is: {model_dir}')
    print(f'layoutreader_model_dir is: {layoutreader_model_dir}')

    json_url = 'https://github.com/opendatalab/MinerU/raw/master/magic-pdf.template.json'
    config_file_name = 'magic-pdf.json'
    home_dir = os.path.expanduser('~')
    config_file = os.path.join(home_dir, config_file_name)

    json_mods = {
        'models-dir': model_dir,
        'layoutreader-model-dir': layoutreader_model_dir,
    }

    download_and_modify_json(json_url, config_file, json_mods)
    print(f'The configuration file has been configured successfully, the path is: {config_file}')

遇到报错说模型不存在：

修改为’zxyayase/layoutreader’即可

3.验证json文件

如果 JSON 中不存在以下项目，请手动添加必填项目并删注释内容。

{
    // other config
    "layout-config": {
        "model": "doclayout_yolo" // Please change to "layoutlmv3" when using layoutlmv3.
    },
    "formula-config": {
        "mfd_model": "yolo_v8_mfd",
        "mfr_model": "unimernet_small",
        "enable": true  // The formula recognition feature is enabled by default. If you need to disable it, please change the value here to "false".
    },
    "table-config": {
        "model": "rapid_table",  // Default to using "rapid_table", can be switched to "tablemaster" or "struct_eqtable".
        "sub_model": "slanet_plus",  // When the model is "rapid_table", you can choose a sub_model. The options are "slanet_plus" and "unitable"
        "enable": true, // The table recognition feature is enabled by default. If you need to disable it, please change the value here to "false".
        "max_time": 400
    }
}

4.cpu运行

wget https://gcore.jsdelivr.net/gh/opendatalab/MinerU@master/demo/small_ocr.pdf
magic-pdf -p small_ocr.pdf -o ./output

可见每页的处理时间大概是20多s。

5. gpu运行

修改【用户目录】中配置文件 magic-pdf.json 中”device-mode”的值

{
  "device-mode":"cuda"
}

再次执行

magic-pdf -p small_ocr.pdf -o ./output