Update app.py

app.py

@@ -1,6 +1,6 @@
 # app.py
 import json
-import requests
+import requests, threading
 import sseclient
 import gradio as gr
 from server import setup_mixinputs, launch_vllm_server
@@ -81,5 +81,5 @@ with gr.Blocks(title="🎨 Mixture of Inputs (MoI) Demo") as demo:
 
 if __name__ == "__main__":
     setup_mixinputs()
-    launch_vllm_server(beta=1.0)
+    threading.Thread(target=launch_vllm_server, daemon=True).start()
     demo.launch()