que

app.py

@@ -103,15 +103,16 @@ with gr.Blocks(title="Live Coding HR Assistant") as demo:
         fn=generate_response,
         inputs=[task_description, code_snippet, request_type, mode],
         outputs=output,
-        api_name="predict"
+        api_name="predict",
+        concurrency_limit=1  # Set concurrency limit here
     )
 
 # Configure queue and launch
 demo.queue(
-    concurrency_count=1,
-    max_size=10
+    max_size=10  # Removed concurrency_count
 ).launch(
     server_name="0.0.0.0",
     server_port=7860,
     share=True
+    # You can optionally add max_threads here if needed, e.g., max_threads=10
 )