diff --git a/examples/seq2seq/task_question_answer_generation_by_seq2seq.py b/examples/seq2seq/task_question_answer_generation_by_seq2seq.py
index bb43c7d8..3e4a4718 100644
--- a/examples/seq2seq/task_question_answer_generation_by_seq2seq.py
+++ b/examples/seq2seq/task_question_answer_generation_by_seq2seq.py
@@ -145,7 +145,7 @@ def predict(self, inputs, output_ids, states):
     def generate(self, passage, topk=1, topp=0.95):
         token_ids, segment_ids = tokenizer.encode(passage, maxlen=max_p_len)
         a_ids = self.random_sample([token_ids, segment_ids], n=1, topp=topp)[0]  # 基于随机采样
-        token_ids += list(a_ids)
+        token_ids += list(a_ids.cpu().numpy())
         segment_ids += [1] * len(a_ids)
         q_ids = self.beam_search([token_ids, segment_ids], topk=topk)[0]  # 基于beam search
         return (tokenizer.decode(q_ids.cpu().numpy()), tokenizer.decode(a_ids.cpu().numpy()))