Finish loading in pyharp's worker before process_fn can reserve a GPU.

app.py

@@ -65,7 +65,15 @@ def load_model():
         model_loading = False
 
 
-threading.Thread(target=load_model, daemon=True).start()
+_loader = threading.Thread(target=load_model, daemon=True)
+_loader.start()
+
+# pyharp's worker imports this file to reach process_fn (as __mp_main__ rather than
+# __main__), so it finishes loading here, before process_fn can reserve a GPU, instead
+# of racing the check there. The server itself keeps loading in the background, which
+# lets it start at once.
+if __name__ != "__main__":
+    _loader.join()
 
 
 model_card = ModelCard(