Clerk2.5/pdf_converter_v2/config.py

# Copyright (c) Opendatalab. All rights reserved.

"""
配置文件 v2 - 从配置文件读取配置（不使用环境变量）
"""

import os
from .config_loader import get_config_loader
from .utils.device_env import detect_device_kind

# 加载配置文件
_config = get_config_loader()

# 设备环境：nvi（NVIDIA GPU）/ npu（华为昇腾 NPU）/ cpu，用于按环境选择 VLLM_PLUGINS、PADDLE_OCR_DEVICE 等
# 优先从配置文件读取，如果未配置则自动检测
_device_kind_from_config = _config.get_str("device_kind", "")
DEVICE_KIND = _device_kind_from_config if _device_kind_from_config else detect_device_kind()

# 默认模型配置
DEFAULT_MODEL_NAME = "OpenDataLab/MinerU2.5-2509-1.2B"
DEFAULT_GPU_MEMORY_UTILIZATION = _config.get_float("default_gpu_memory_utilization", 0.9)
DEFAULT_DPI = _config.get_int("default_dpi", 200)
DEFAULT_MAX_PAGES = _config.get_int("default_max_pages", 10)

# v2 特有配置（外部API相关）
DEFAULT_API_URL = _config.get_str("api_url", "http://127.0.0.1:5282")

# 根据设备类型自动选择 MinerU backend（配置文件显式指定时优先）
# npu  -> vlm-http-client（NPU 不支持本地 vllm，需通过 http 调用外部推理服务）
# nvi  -> vlm-auto-engine（NVIDIA GPU 本地推理）
# cpu  -> pipeline（纯 CPU 通用管线）
def _auto_backend() -> str:
    if DEVICE_KIND == "npu":
        return "vlm-http-client"
    elif DEVICE_KIND == "nvi":
        return "vlm-auto-engine"
    return "pipeline"

def _auto_server_url() -> str:
    """server_url 仅在 http-client 类 backend 时有意义"""
    if DEVICE_KIND == "npu":
        return "http://172.17.0.1:30000"
    return ""

_backend_from_config = _config.get_str("backend", "")
DEFAULT_BACKEND = _backend_from_config if _backend_from_config else _auto_backend()

_server_url_from_config = _config.get_str("server_url", "")
DEFAULT_SERVER_URL = _server_url_from_config if _server_url_from_config else _auto_server_url()

DEFAULT_PARSE_METHOD = _config.get_str("parse_method", "auto")
DEFAULT_START_PAGE_ID = _config.get_int("start_page_id", 0)
DEFAULT_END_PAGE_ID = _config.get_int("end_page_id", -1)
DEFAULT_LANGUAGE = _config.get_str("language", "ch")
DEFAULT_RESPONSE_FORMAT_ZIP = _config.get_bool("response_format_zip", True)
DEFAULT_RETURN_MIDDLE_JSON = _config.get_bool("return_middle_json", True)
DEFAULT_RETURN_MODEL_OUTPUT = _config.get_bool("return_model_output", False)
DEFAULT_RETURN_MD = _config.get_bool("return_md", True)
DEFAULT_RETURN_IMAGES = _config.get_bool("return_images", True)
DEFAULT_RETURN_CONTENT_LIST = _config.get_bool("return_content_list", True)
DEFAULT_TABLE_ENABLE = _config.get_bool("table_enable", True)
DEFAULT_FORMULA_ENABLE = _config.get_bool("formula_enable", True)

# PaddleOCR 配置
PADDLEOCR_CMD = _config.get_str("paddleocr_cmd", "paddleocr")
PADDLE_DOC_PARSER_CMD = _config.get_str("paddle_doc_parser_cmd", "paddleocr")

# PaddleOCR 设备参数（留空则根据 DEVICE_KIND 自动选择）
# npu -> "npu:0" | nvi -> ""（PaddlePaddle 自动使用 CUDA） | cpu -> "cpu"
def _auto_paddle_device() -> str:
    if DEVICE_KIND == "npu":
        return "npu:0"
    elif DEVICE_KIND == "cpu":
        return "cpu"
    return ""  # nvi: PaddlePaddle 自动检测 GPU

_paddle_device_from_config = _config.get_str("paddle_ocr_device", "")
PADDLE_OCR_DEVICE = _paddle_device_from_config if _paddle_device_from_config else _auto_paddle_device()
PADDLE_OCR_DEVICES = _config.get_str("paddle_ocr_devices", "")

# VL 识别后端（仅 NPU 环境需要通过 vllm-server 调用外部推理，NVIDIA 本地推理不需要）
def _auto_vl_rec_backend() -> str:
    if DEVICE_KIND == "npu":
        return "vllm-server"
    return ""  # nvi/cpu: 不使用 VL 识别后端

def _auto_vl_rec_server_url() -> str:
    if DEVICE_KIND == "npu":
        return "http://172.17.0.1:8118/v1"
    return ""

_vl_backend_from_config = _config.get_str("vl_rec_backend", "")
VL_REC_BACKEND = _vl_backend_from_config if _vl_backend_from_config else _auto_vl_rec_backend()

_vl_url_from_config = _config.get_str("vl_rec_server_url", "")
VL_REC_SERVER_URL = _vl_url_from_config if _vl_url_from_config else _auto_vl_rec_server_url()

# 日志配置（可选）
LOG_DIR = _config.get_str("log_dir", "./logs")
LOG_LEVEL = _config.get_str("log_level", "INFO")
LOG_TO_FILE = _config.get_bool("log_to_file", True)
LOG_TO_CONSOLE = _config.get_bool("log_to_console", True)