Update app.py

app.py

@@ -1,7 +1,7 @@
 ##########====================================================================################
 ##########====================PRODUCTION VERSION -- vLLM, GRADIO=====================###########
 ##########====================================================================################
-import os, subprocess, threading, time
+import os, time
 import requests
 from typing import List, Dict, Tuple
 from datetime import datetime
@@ -121,7 +121,6 @@ def launch_claude_api(sys_msg, user_msg):
 
 def generate_mockup_data(category, num_data_rows, columns, a_instruction,
                          progress=gr.Progress()):
-    wait_for_vllm_ready()
     progress(0.2, desc="Generating...")
     msg = invoke_messages(
         rows_num=int(num_data_rows or 10),
@@ -236,9 +235,11 @@ def render_interface():
             inputs=[category, out],
             outputs=btn_json
         )
+    return demo
 
 
 if __name__ == "__main__":
+    wait_for_vllm_ready()
     app = render_interface()
     app.queue(default_concurrency_limit=1)
     app.launch(server_name="0.0.0.0", server_port=7860)