update generate parameters

app.py

@@ -38,8 +38,7 @@ def chat_func(input_utterance: str, history: Optional[List[str]] = None):
 
     output_ids = model.generate(input_ids,
                                 max_new_tokens=30,
-                                top_p=0.95,
-                                do_sample=True,
+                                top_k=32,
                                 num_beams=4,
                                 repetition_penalty=1.2,
                                 no_repeat_ngram_size=4)[0]