fix(ocr): run warm-up in background so server accepts connections immediately

main.py

@@ -9,6 +9,7 @@ Usage:
   uvicorn main:app --host 0.0.0.0 --port 8000 --reload
 """
 
+import asyncio
 import logging
 from contextlib import asynccontextmanager
 
@@ -26,20 +27,26 @@ logger = logging.getLogger(__name__)
 _reader: PaddleOCRReader | None = None
 
 
+async def _background_warmup():
+    import numpy as np
+    logger.info("Background warm-up starting...")
+    try:
+        await asyncio.get_event_loop().run_in_executor(
+            None, lambda: _reader.extract(np.ones((200, 600, 3), dtype=np.uint8) * 255)
+        )
+        logger.info("Background warm-up complete.")
+    except Exception as exc:
+        logger.warning("Background warm-up failed (non-fatal): %s", exc)
+
+
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     global _reader
     _reader = PaddleOCRReader()
-    # Warm up PaddlePaddle JIT compilation so the first real request is fast.
-    # Without this, the first inference takes 60-90s (JIT compile on CPU),
-    # which exceeds the app's 45s timeout even though the server is "ready".
-    import numpy as np
-    logger.info("Warming up OCR inference engine...")
-    try:
-        _reader.extract(np.ones((200, 600, 3), dtype=np.uint8) * 255)
-        logger.info("OCR warm-up complete.")
-    except Exception as exc:
-        logger.warning("OCR warm-up failed (non-fatal): %s", exc)
+    # Run warm-up in background so the server starts accepting connections
+    # immediately. First real request after restart may still be slow (JIT),
+    # but the server won't appear unresponsive during startup.
+    asyncio.create_task(_background_warmup())
     yield
     _reader = None