[Jammy123-prog]

Problem

1. Slow startup: MainWindow.__init__ builds the full PPStructureV3 pipeline unconditionally. That loads ~14 bundled models (~1.1 GB: PP-DocLayout-L, SLANeXt/SLANet, RT-DETR-L cell detectors, etc.) before the main window appears — on a machine without a warm model cache this adds several minutes to launch, even for users who only annotate plain text and never use table recognition. The eager warm-up call table_ocr.predict("./data/paddle.png") has the same cost.
2. Startup crash with PaddleX >= 3.7: the HPI auto-tuner may pick the oneDNN backend for the table models, which aborts with:

NotImplementedError: (Unimplemented) ConvertPirAttribute2RuntimeAttribute not support
[pir::ArrayAttribute<pir::DoubleAttribute>]
(at ..\paddle\fluid\framework\new_executor\instruction\onednn\onednn_instruction.cc:118)

Observed with paddlepaddle-gpu 3.3.1 + paddlex 3.7.2 on Windows 11 (RTX 1000 Ada, CUDA 12.6).

Fix

• Initialize self.table_ocr = None and construct the pipeline lazily in a new _get_table_ocr() helper, called from TableRecognition(). Table features work exactly as before — models simply load on first use.
• Drop the eager table warm-up call (the text-OCR warm-up is kept).
• Pass enable_mkldnn=False, enable_hpi=False when constructing PPStructureV3 so a stable, working backend is used (mirrors what PPOCRLabel already does for its PaddleOCR instance).
• The language-switch path now just resets self.table_ocr = None; the pipeline is rebuilt on demand with the new language.

Result

Startup now only loads det/rec models (Creating model lines drop from 16 to 2 in a typical session); resident GPU memory drops accordingly, which also helps low-VRAM laptops. Table recognition remains fully functional, loading its models on first use (~1 min one-time cost, then cached).

Verified: offscreen launch test with custom --det_model_dir/--rec_model_dir models loads only 2 models and runs clean; table recognition path works after lazy init.

中文说明

问题：① 启动时无条件构建 PPStructureV3 全套表格识别管线（约 14 个模型、1.1GB），纯文本标注场景也要多等数分钟；② PaddleX ≥ 3.7 的 HPI 自动调优可能为表格模型选中 oneDNN 后端，启动即抛 NotImplementedError 崩溃。

修复：表格管线改为首次使用时懒加载（新增 _get_table_ocr()）；删除启动阶段的表格推理预热；构造时显式 enable_mkldnn=False, enable_hpi=False（与 PPOCRLabel 对 PaddleOCR 的处理一致）；切换语言处仅置空 table_ocr，按需重建。启动只加载检测/识别两个模型，表格标注功能不受影响。已在 Windows 11 + paddlepaddle-gpu 3.3.1 环境验证。