MinerU 3.0
从 PDF 到 Markdown 的三后端之路

📦 依赖矩阵(按后端切分)

# mineru/backend/vlm/vlm_analyze.py:80-102
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor

model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_path,
    device_map={"": device},
    dtype="auto"
)
processor = AutoProcessor.from_pretrained(model_path, use_fast=True)

🚀 四种推理框架并存

🔄 推理主流程

⚙️ Batch Size 自适应

# mineru/backend/vlm/utils.py:94-110
def set_default_batch_size() -> int:
    gpu_memory = get_vram(device)
    if gpu_memory >= 16:
        batch_size = 8
    elif gpu_memory >= 8:
        batch_size = 4
    else:
        batch_size = 1

🧩 核心类 BatchAnalyze

🏷️ Magic Model: 标签 → BlockType 映射

# mineru/backend/pipeline/pipeline_magic_model.py:18-42
PP_DOCLAYOUT_V2_LABELS_TO_BLOCK_TYPES = {
    "image": BlockType.IMAGE,
    "table": BlockType.TABLE,
    "display_formula": BlockType.INTERLINE_EQUATION,
    "text": BlockType.TEXT,
    # ...
}

VISUAL_MAIN_TYPES  = (BlockType.IMAGE, BlockType.TABLE, BlockType.CHART, BlockType.CODE)
VISUAL_CHILD_TYPES = (BlockType.CAPTION, BlockType.FOOTNOTE)

# mineru/backend/hybrid/hybrid_analyze.py:1-150
def hybrid_analyze(pdf_bytes, lang_list, parse_method="auto", ...):
    # 1. 判断是否需要 OCR(行 50-58)
    _ocr_enable = ocr_classify(pdf_bytes, parse_method)

    # 2. 需要 OCR 则走 Pipeline OCR
    if _ocr_enable:
        ocr_res_list = ocr_det(...)

    # 3. 表格/公式等难块交给 VLM 二次验证
    # 4. 最后融合中间 JSON

# mineru/backend/office/docx_analyze.py:11-29
def office_docx_analyze(file_bytes, image_writer=None):
    file_stream = BytesIO(file_bytes)
    results = convert_binary(file_stream)   # DocxConverter
    middle_json = result_to_middle_json(results, image_writer)
    return middle_json, results

# 统一中间表示
middle_json = {
    "meta_info": {...},
    "doc_title": str,
    "doc_layout_result": [...],
    "para_blocks": [
        {
            "type": BlockType,
            "blocks": [
                {
                    "type": BlockType,
                    "lines": [
                        {"spans": [
                            {
                                "type": ContentType,
                                "content": str,
                                "bbox": [x1, y1, x2, y2],
                            }
                        ]}
                    ]
                }
            ]
        }
    ]
}

🌐 多语言智能换行

# mineru/backend/vlm/vlm_middle_json_mkcontent.py:58-90
block_lang = detect_lang(block_text)

if block_lang in {'zh', 'ja', 'ko'}:  # CJK:换行不加空格
    para_text += content
else:                                    # 西文:处理连字符
    if is_hyphen_at_line_end(content):
        para_text += content[:-1]     # 删掉行尾连字符
    else:
        para_text += f"{content} "

📐 LaTeX 定界符可配置

delimiters = {
    'display': {'left': '$$', 'right': '$$'},   # 行间公式
    'inline':  {'left': '$',  'right': '$'}     # 行内公式
}
# 可通过 config.yaml 改成 \[...\] / \(...\)

📋 ParseRequestOptions(REST 参数全景)

# mineru/cli/fast_api.py:130-149
@dataclass
class ParseRequestOptions:
    files: list[UploadFile]
    lang_list: list[str]
    backend: str              # "vlm" / "pipeline" / "hybrid-ocr"
    parse_method: str         # "auto" / "txt" / "ocr"
    formula_enable: bool
    table_enable: bool
    server_url: Optional[str]   # 远程 VLM 服务器
    return_md: bool
    return_middle_json: bool
    return_model_output: bool
    return_content_list: bool
    return_images: bool
    response_format_zip: bool