Update app.py

app.py

@@ -32,7 +32,7 @@ def respond(message, history, system_message, max_tokens, temperature, top_p):
 
     # Генерация ответа
     output_ids = model.generate(
-        input_ids.to('cuda'),
+        input_ids.to('cpu'),
         max_new_tokens=max_tokens,
         temperature=temperature,
         top_p=top_p