OCR Worker 改用 RapidOCR/ONNX，修复 Paddle SIGILL。

deploy/ocr-worker/app.py

@@ -1,4 +1,4 @@
-"""局域网 OCR 服务：在带 NVIDIA 显卡的机器上运行，供成绩档案系统调用。"""
+"""局域网 OCR 服务：RapidOCR(ONNX)，不依赖 Paddle，避免 SIGILL/cuDNN 问题。"""
 
 import logging
 import os
@@ -9,19 +9,14 @@ from pathlib import Path
 from fastapi import FastAPI, File, Header, HTTPException, UploadFile
 from PIL import Image
 
-os.environ.setdefault("OPENCV_IO_ENABLE_OPENEXR", "0")
-os.environ.setdefault("FLAGS_use_mkldnn", "0")
-
 logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
 logger = logging.getLogger("ocr-worker")
 
 OCR_MAX_SIDE = int(os.getenv("OCR_MAX_SIDE", "1280"))
 OCR_API_KEY = os.getenv("OCR_API_KEY", "").strip()
-OCR_USE_GPU = os.getenv("OCR_USE_GPU", "true").lower() in {"1", "true", "yes"}
 
-app = FastAPI(title="Grade Archive OCR Worker", version="1.0.0")
+app = FastAPI(title="Grade Archive OCR Worker", version="2.0.0")
 _engine = None
-_engine_mode = "none"
 
 
 def _check_key(key: str | None) -> None:
@@ -29,54 +24,15 @@ def _check_key(key: str | None) -> None:
         raise HTTPException(status_code=401, detail="Invalid OCR API key")
 
 
-def _create_engine(use_gpu: bool):
-    from paddleocr import PaddleOCR
+def get_engine():
+    global _engine
+    if _engine is None:
+        from rapidocr_onnxruntime import RapidOCR
 
-    return PaddleOCR(
-        use_angle_cls=False,
-        lang="ch",
-        show_log=False,
-        use_gpu=use_gpu,
-        enable_mkldnn=False,
-        ir_optim=False,
-        det_limit_side_len=min(OCR_MAX_SIDE, 1280),
-        rec_batch_num=8,
-    )
-
-
-def get_engine(force_cpu: bool = False):
-    global _engine, _engine_mode
-    if _engine is not None and not force_cpu:
-        return _engine
-
-    modes: list[bool] = [False] if force_cpu or not OCR_USE_GPU else [True, False]
-    last_err: Exception | None = None
-    for use_gpu in modes:
-        try:
-            logger.info("Loading PaddleOCR use_gpu=%s", use_gpu)
-            _engine = _create_engine(use_gpu)
-            _engine_mode = "gpu" if use_gpu else "cpu"
-            logger.info("PaddleOCR ready mode=%s", _engine_mode)
-            return _engine
-        except Exception as exc:
-            last_err = exc
-            logger.warning("PaddleOCR init failed use_gpu=%s: %s", use_gpu, exc)
-            _engine = None
-            _engine_mode = "none"
-
-    hint = ""
-    err_text = str(last_err or "")
-    if "libGL" in err_text:
-        hint = " 请执行: sudo bash deploy/install-ocr-deps.sh 后重启 OCR"
-    elif any(x in err_text.lower() for x in ("cuda", "cudnn", "gpu", "out of memory")):
-        hint = " 显存不足或 CUDA 异常，可设置 OCR_USE_GPU=false 用 CPU"
-    raise RuntimeError(f"PaddleOCR 初始化失败: {last_err}{hint}") from last_err
-
-
-def _reset_engine():
-    global _engine, _engine_mode
-    _engine = None
-    _engine_mode = "none"
+        logger.info("Loading RapidOCR (ONNX CPU)…")
+        _engine = RapidOCR()
+        logger.info("RapidOCR ready")
+    return _engine
 
 
 def _bbox_from_box(box: list) -> list[float]:
@@ -110,32 +66,31 @@ def _prepare_image_bytes(content: bytes) -> tuple[bytes, float, float, int, int]
         return buf.getvalue(), scale_x, scale_y, orig_w, orig_h
 
 
-def _run_ocr_impl(content: bytes) -> dict:
+def run_ocr_on_bytes(content: bytes) -> dict:
     engine = get_engine()
     image_bytes, scale_x, scale_y, orig_w, orig_h = _prepare_image_bytes(content)
     with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmp:
         tmp.write(image_bytes)
         tmp_path = tmp.name
     try:
-        result = engine.ocr(tmp_path, cls=False)
+        result, _elapsed = engine(tmp_path)
     finally:
         Path(tmp_path).unlink(missing_ok=True)
 
     lines: list[dict] = []
-    if result and result[0]:
-        for item in result[0]:
+    if result:
+        for item in result:
             if not item or len(item) < 2:
                 continue
-            box, rec = item[0], item[1]
-            text = rec[0] if rec else ""
-            conf = float(rec[1]) if rec and len(rec) > 1 else 0.0
+            box, text = item[0], item[1]
+            conf = float(item[2]) if len(item) > 2 else 0.0
             if not text:
                 continue
             if scale_x != 1.0 or scale_y != 1.0:
                 box = _scale_box(box, scale_x, scale_y)
             lines.append(
                 {
-                    "text": text,
+                    "text": str(text),
                     "confidence": conf,
                     "box": box,
                     "bbox": _bbox_from_box(box),
@@ -147,17 +102,13 @@ def _run_ocr_impl(content: bytes) -> dict:
         "lines": lines,
         "width": orig_w,
         "height": orig_h,
-        "engine_mode": _engine_mode,
+        "engine_mode": "rapidocr-onnx",
     }
 
 
-def run_ocr_on_bytes(content: bytes) -> dict:
-    return _run_ocr_impl(content)
-
-
 @app.get("/health")
 def health():
-    return {"status": "ok", "gpu_requested": OCR_USE_GPU, "engine_mode": _engine_mode}
+    return {"status": "ok", "engine": "rapidocr-onnxruntime"}
 
 
 @app.post("/api/ocr/regions")