Sync CUDA before timing for accurate runtime

app/safety.py

@@ -7,6 +7,7 @@ import time
 
 import cv2
 import numpy as np
+import torch
 from PIL import Image
 
 from .config import IMAGE_EXTS
@@ -405,6 +406,11 @@ class SafetyAnalyzer:
             seg_mask_union=seg_mask_union,
             hazard_mask=hazard_mask,
         )
+        try:
+            if torch.cuda.is_available():
+                torch.cuda.synchronize()
+        except Exception:
+            pass
         runtime_ms = (time.perf_counter() - t0) * 1000.0
         safe_area_pct = float(safe_display_mask.mean()) * 100.0
         hazard_pct = 100.0 - safe_area_pct