perf: MAX_BASE_TILES=4 + 50s time budget to prevent timeout

Dockerfile

@@ -15,7 +15,7 @@ RUN pip install --no-cache-dir -r requirements.txt
 COPY api/ ./api/
 
 # HF Space: limit tiles for CPU inference speed
-ENV MAX_BASE_TILES=8
+ENV MAX_BASE_TILES=4
 ENV PORT=7860
 EXPOSE 7860
 

api/scanner.py

@@ -11,6 +11,7 @@ Revenue estimates use 2025 Korean market rates:
 """
 import math
 import os
+import time
 import numpy as np
 import requests
 from collections import deque
@@ -299,8 +300,17 @@ def scan_region(
                 batch_keys.append((vx, vy))
                 batch_feats.append(_img_features(crop_256))
 
+    # 시간 예산 50s: 초과 시 남은 타일은 prob=0으로 처리 (타임아웃 방지)
+    _BUDGET_S = 50.0
+    _t0 = time.monotonic()
     if batch_imgs:
-        probs = infer_batch(batch_imgs)
+        probs: list[float] = []
+        mini = 8
+        for i in range(0, len(batch_imgs), mini):
+            if time.monotonic() - _t0 > _BUDGET_S:
+                probs.extend([0.0] * (len(batch_imgs) - len(probs)))
+                break
+            probs.extend(infer_batch(batch_imgs[i : i + mini]))
         for (vx, vy), prob, feat in zip(batch_keys, probs, batch_feats):
             tile_data[(vx, vy)] = (prob, feat)