diff --git a/python/paddle/fluid/tests/unittests/dist_transformer.py b/python/paddle/fluid/tests/unittests/dist_transformer.py
index ab44954811562b8f74e368a551e855948f90af87..23abd7953f96735b447aa9edde62a848db5367fb 100644
--- a/python/paddle/fluid/tests/unittests/dist_transformer.py
+++ b/python/paddle/fluid/tests/unittests/dist_transformer.py
@@ -1159,6 +1159,7 @@ def prepare_encoder(src_word,
             name=pos_enc_param_name,
             trainable=False,
             initializer=fluid.initializer.ConstantInitializer(0.001)))
+    str_pos_enc.stop_gradient = True
     enc_input = src_word_emb + src_pos_enc
     return layers.dropout(
         enc_input,