Update app.py

app.py

@@ -33,7 +33,7 @@ else:
 # Dispatch Errors
 
 
-@spaces.GPU
+@spaces.GPU(duration=150)
 def chat(message, history, temperature, top_p, top_k, max_tokens):
     start_time = time.time()
     chat = []
@@ -72,8 +72,7 @@ def chat(message, history, temperature, top_p, top_k, max_tokens):
     tokens_per_second = tokens / total_time if total_time > 0 else 0
 
     timing_info = f"\n\nTime taken to first token: {first_token_time:.2f} seconds\nTokens per second: {tokens_per_second:.2f}"
-    gr.Info(timing_info)
-    yield partial_text 
+    yield partial_text +  timing_info
 
 
 demo = gr.ChatInterface(