Reopen test_parallel_executor_transformer (#12452)

* reopen test_parallel_executor_transformer

* save the CPU runing time

python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py

@@ -167,10 +167,9 @@ class TestTransformer(TestParallelExecutorBase):
                     writer.append_tensor(t)
                 writer.complete_append_tensor()
 
-    @unittest.skip("transformer is buggy in multi gpu")
     def test_main(self):
         self.check_network_convergence(transformer, use_cuda=True)
-        self.check_network_convergence(transformer, use_cuda=False)
+        self.check_network_convergence(transformer, use_cuda=False, iter=5)
 
 
 if __name__ == '__main__':