diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
index 1f47d87811cf4ca63bda63da860e2ac3b9de1e7e..2d1e0e98498af5e943b23797888c55f3df4991e0 100644
--- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
+++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py
@@ -25,6 +25,8 @@ import paddle.dataset.wmt16 as wmt16
 import os
 from feed_data_reader import FeedDataReader
 
+os.environ['CPU_NUM'] = str(4)
+
 
 class ModelHyperParams(object):
     # Dictionary size for source and target language. This model directly uses
@@ -185,10 +187,6 @@ def get_feed_data_reader():
 
 
 class TestTransformer(TestParallelExecutorBase):
-    @classmethod
-    def setUpClass(cls):
-        os.environ['CPU_NUM'] = str(4)
-
     def test_main(self):
         if core.is_compiled_with_cuda():
             self.check_network_convergence(