Update app.py

app.py

@@ -116,7 +116,6 @@ def stream_chat(message: str, history: list, model: str, temperature: float, max
         response = client.chat(
             model=model,
             messages=conversation,
-            stream=True,
             options={
                 'num_predict': max_new_tokens,
                 'temperature': temperature,
@@ -128,7 +127,7 @@ def stream_chat(message: str, history: list, model: str, temperature: float, max
             },
         )
 
-        return response
+        yield response
 
 
 
@@ -142,6 +141,7 @@ def main(message: str, history: list, model: str, temperature: float, max_new_to
         else:
             if not process:
                 launch()
+                
             response = stream_chat(
                 message,
                 history,
@@ -153,6 +153,8 @@ def main(message: str, history: list, model: str, temperature: float, max_new_to
                 penalty             
             )
             
+            print(response)
+            
             buffer = ""
             for chunk in response:
                 buffer += chunk["message"]["content"]