Update app.py

app.py

@@ -128,10 +128,7 @@ def stream_chat(message: str, history: list, model: str, temperature: float, max
             },
         )
 
-        buffer = ""
-        for chunk in response:
-            buffer += chunk["message"]["content"]
-            yield buffer
+        return response
 
 
 
@@ -145,7 +142,7 @@ def main(message: str, history: list, model: str, temperature: float, max_new_to
         else:
             if not process:
                 launch()
-            stream_chat(
+            response = stream_chat(
                 message,
                 history,
                 model,
@@ -155,6 +152,11 @@ def main(message: str, history: list, model: str, temperature: float, max_new_to
                 top_k,
                 penalty             
             )
+            
+            buffer = ""
+            for chunk in response:
+                buffer += chunk["message"]["content"]
+                yield buffer
 
 
 chatbot = gr.Chatbot(height=600, placeholder=DESCRIPTION)