Update app.py

app.py

@@ -116,7 +116,6 @@ def stream_chat(message: str, history: list, model: str, temperature: float, max
             model=model,
             messages=conversation,
             keep_alive="60s",
-            stream=True,
             options={
                 'num_predict': max_new_tokens,
                 'temperature': temperature,
@@ -127,10 +126,8 @@ def stream_chat(message: str, history: list, model: str, temperature: float, max
             },
         )
         
-        terminate()
-        
         print(response)
-        yield response['message']['content']
+        return response['message']['content']
 
 
 
@@ -156,10 +153,8 @@ def main(message: str, history: list, model: str, temperature: float, max_new_to
                 penalty             
             )
             
-            buffer = ""
-            for chunk in response:
-                buffer += chunk
-                yield buffer
+            terminate()
+            yield response