Register ZeroGPU @spaces.GPU tasks at startup

app.py

@@ -15,6 +15,17 @@ from worlds import get_world, WORLDS, list_worlds
 
 MOCK = os.environ.get("TINYWORLD_MOCK", "0") == "1"
 
+# On a ZeroGPU Space (TINYWORLD_INFER=local) the @spaces.GPU functions must be
+# registered at startup or ZeroGPU shuts the Space down. Import inference eagerly
+# in that mode only (loads torch but NOT the models — those stay lazy). Skipped on
+# mock/modal so dev and tests never need torch/spaces.
+if not MOCK and os.environ.get("TINYWORLD_INFER", "modal").lower() == "local":
+    try:
+        import inference  # noqa: F401  (registers ZeroGPU GPU tasks)
+        print("[app] ZeroGPU inference module loaded")
+    except Exception as e:
+        print(f"[app] inference preimport failed: {e}")
+
 MOOD_EMOJI = {
     "happy": "😊", "stressed": "😰", "bored": "😐", "excited": "🤩",
     "hungry": "🍕", "tired": "😴", "nostalgic": "🥹", "curious": "🧐",