PPStructureV3 loads GeneralOCR (det + rec models) twice when use_table_recognition=True

[Samdo3]

Checklist:

• [✓ ] 查找历史相关issue寻求解答
• [✓ ] 翻阅FAQ
• [✓ ] 翻阅PaddleX 文档
• [ ✓] 确认bug是否在新版本里还未修复

描述问题

When using PPStructureV3 with lang='korean' and use_table_recognition=True(the default), the GeneralOCR sub-pipeline (text detection + text recognition models) is loaded into memory twice — once by the parent _LayoutParsingPipelineV2 at init, and once by the child _TableRecognitionPipelineV2 on the first predict() call.

In our case, lang='korean' auto-selects korean_PP-OCRv5_mobile_rec as the recognition model (instead of the default Chinese PP-OCRv5_server_rec). Both PP-OCRv5_server_det and korean_PP-OCRv5_mobile_rec are loaded twice, unnecessarily doubling GPU/CPU memory consumption. This issue likely affects all non-default language configurations, not just Korean.

Related issue

• PaddleOCR#17266: Under certain conditions, general_ocr_pipeline remains None inside TableRecognition, causing AttributeError: 'NoneType' object has no attribute 'text_rec_model'. Same root cause (TableRecognition does not share parent's GeneralOCR), different symptom (crash vs memory waste).

复现

1. 您是否已经正常运行我们提供的教程？

Yes. The PPStructureV3 tutorial runs successfully.

2. 您是否在教程的基础上修改代码内容？还请您提供运行的代码

"""
Test script to reproduce duplicate GeneralOCR model loading in PPStructureV3.

Reproduces the exact configuration used in our project:
  - backend/app/core/ocr_engine.py (KoreanPPOCRv5Engine)
  - lang='korean' → auto-selects korean_PP-OCRv5_mobile_rec
  - All unnecessary features disabled
"""

from paddleocr import PPStructureV3
import numpy as np

print("=" * 60)
print("Initializing PPStructureV3 with Korean PP-OCRv5 model...")
print("  - lang: korean (auto-selects korean_PP-OCRv5_mobile_rec)")
print("  - Table recognition: Enabled")
print("  - Formula recognition: Disabled")
print("  - Region detection: Disabled")
print("=" * 60)

pipeline = PPStructureV3(
    lang='korean',
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False,
    use_table_recognition=True,
    use_formula_recognition=False,
    use_region_detection=False,
    format_block_content=True,
)

print("=" * 60)
print("Initialization done. Running predict()...")
print("=" * 60)

test_img = np.ones((100, 300, 3), dtype=np.uint8) * 255
try:
    result = list(pipeline.predict(
        test_img,
        use_table_recognition=True,
        format_block_content=True,
    ))
    print("Predict done.")
except Exception as e:
    print(f"Predict error: {e}")

print("=" * 60)
print("Test complete.")
print("=" * 60)

3. 您使用的数据集是？

Not dataset-dependent. Any input triggers this issue since the duplicate loading occurs on the first predict() call regardless of input content.

4. 请提供您出现的报错信息及相关log

No error — the pipeline works correctly, but the log shows det + rec models loaded twice. The duplicate 4 models are lazily loaded on the first predict() call (after "Initialization done" in the log):

============================================================
Initializing PPStructureV3 with Korean PP-OCRv5 model...
  - lang: korean (auto-selects korean_PP-OCRv5_mobile_rec)
  - Table recognition: Enabled
  - Formula recognition: Disabled
  - Region detection: Disabled
============================================================
Creating model: ('PP-DocLayout_plus-L', None)
Creating model: ('PP-OCRv5_server_det', None)              ← 1st load
Creating model: ('korean_PP-OCRv5_mobile_rec', None)       ← 1st load
Creating model: ('PP-LCNet_x1_0_table_cls', None)
Creating model: ('SLANeXt_wired', None)
Creating model: ('SLANet_plus', None)
Creating model: ('RT-DETR-L_wired_table_cell_det', None)
Creating model: ('RT-DETR-L_wireless_table_cell_det', None)
Creating model: ('PP-Chart2Table', None)
============================================================
Initialization done. Running predict()...
============================================================
Creating model: ('PP-LCNet_x1_0_doc_ori', None)            ← DUPLICATE (disabled by user!)
Creating model: ('PP-LCNet_x1_0_textline_ori', None)       ← DUPLICATE (disabled by user!)
Creating model: ('PP-OCRv5_server_det', None)               ← DUPLICATE (2nd load)
Creating model: ('korean_PP-OCRv5_mobile_rec', None)        ← DUPLICATE (2nd load)
Predict done.
============================================================
Test complete.
============================================================

环境

1. 请提供您使用的PaddlePaddle和PaddleX的版本号

• PaddlePaddle-GPU 3.3.0
• PaddleX 3.4.2
• PaddleOCR 3.4.0

2. 请提供您使用的操作系统信息，如Linux/Windows/MacOS

Windows 11

3. 请问您使用的Python版本是？

Python 3.11.14 (Anaconda)

5. 请问您使用的CUDA/cuDNN的版本号是？

CUDA 12.6 / cuDNN 9.9.0 (compiled), GPU Compute Capability 8.9, Driver 561.17