[refactor] YOLO 객체 탐지 코드 gpu 강제 코드 추가

Author: hhongyeahh

everTale/app/service/yolo_service.py

@@ -9,14 +9,35 @@
 
 YOLO_MODEL_PATH = os.environ["YOLO_MODEL_PATH"]
 
+from ultralytics import YOLO
+import os, torch
+
+def _resolve_yolo_path() -> str:
+    path = os.getenv("YOLO_MODEL_PATH", "/models/my_yolo_model.pt")
+    if not os.path.exists(path):
+        raise FileNotFoundError(f"YOLO model not found at: {path}")
+    return path
+
+def _require_gpu_for_yolo(stage: str = "YOLO load"):
+    if torch.cuda.is_available():
+        return 0  # device index for CUDA
+    # MPS는 Ultralytics 지원이 제한적이므로 필요한 경우만 허용
+    if torch.backends.mps.is_available() and torch.backends.mps.is_built():
+        return "mps"
+    raise RuntimeError(f"[ERROR] No GPU backend during {stage}. CPU is not allowed for YOLO.")
+
 def load_model() -> YOLO:
+    path = _resolve_yolo_path()
+    device = _require_gpu_for_yolo("YOLO load")
     try:
-        model = YOLO(YOLO_MODEL_PATH)
-        print("모델이 성공적으로 로드되었습니다.")
+        model = YOLO(path)
+        # warm-up(선택): 작은 더미로 한 번 실행해 메모리 로딩
+        model.predict(source=np.zeros((64,64,3), dtype=np.uint8), device=device, imgsz=64, verbose=False)
+        print(f"[INFO] YOLO loaded on device={device} from {path}")
         return model
     except Exception as e:
-        print(f"모델 로드 중 오류가 발생했습니다: {e}")
-        return None
+        raise RuntimeError(f"Failed to load YOLO model at {path}: {e}")
+
 
 def _url_to_bgr(url: str) -> np.ndarray:
     resp = requests.get(url, timeout=10)
@@ -35,13 +56,19 @@ def detect_object(image_paths: List[str]) -> Dict[str, Any]:
           탐지 후보가 전혀 없으면 {"index": None, "url": None, "detection": None}
     """
     model = load_model()
+    device = 0 if torch.cuda.is_available() else "mps"  # 위와 일치
     urls = image_paths[:8]
     candidates: List[Dict[str, Any]] = []
 
     for idx, url in enumerate(urls):
         try:
             img = _url_to_bgr(url)
-            results = model.predict(source=img, verbose=False)
+            results = model.predict(
+                source=img,
+                device=device,
+                half=torch.cuda.is_available(),
+                verbose=False
+            )
             if not results or results[0].boxes is None or results[0].boxes.shape[0] == 0:
                 continue