diff --git a/paddle/fluid/operators/batch_norm_op.cu b/paddle/fluid/operators/batch_norm_op.cu
index 8f3691dd15359b42dfacca1c26e2b631a48161b0..36d297ec5523b9e8a136c536165bdb4d3a380c25 100644
--- a/paddle/fluid/operators/batch_norm_op.cu
+++ b/paddle/fluid/operators/batch_norm_op.cu
@@ -75,15 +75,11 @@ class BatchNormKernel<platform::CUDADeviceContext, T>
                  << "CUDNN_BN_MIN_EPSILON instead.";
     }
     epsilon = std::max(epsilon, CUDNN_BN_MIN_EPSILON);
-
-    // TODO(dengkaipeng): use PERSISTENT mode in training may incur errors
-    // in inference period, cuDNN fixed issues on PERSISTENT mode in version
-    // 7.0.2, 7.0.4 and 7.3.0, we disable this mode currently.
-    // #if CUDNN_VERSION_MIN(7, 0, 0)
-    //     mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT;
-    // #else
+#if CUDNN_VERSION_MIN(7, 0, 0)
+    mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT;
+#else
     mode_ = CUDNN_BATCHNORM_SPATIAL;
-    // #endif
+#endif
 
     VLOG(3) << "Setting descriptors.";
     std::vector<int> dims;
@@ -305,15 +301,11 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
                    << "CUDNN_BN_MIN_EPSILON instead.";
       }
       epsilon = std::max(epsilon, CUDNN_BN_MIN_EPSILON);
-
-      // TODO(dengkaipeng): use PERSISTENT mode in training may incur errors
-      // in inference period, cuDNN fixed issues on PERSISTENT mode in version
-      // 7.0.2, 7.0.4 and 7.3.0, we disable this mode currently.
-      // #if CUDNN_VERSION_MIN(7, 0, 0)
-      //       mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT;
-      // #else
+#if CUDNN_VERSION_MIN(7, 0, 0)
+      mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT;
+#else
       mode_ = CUDNN_BATCHNORM_SPATIAL;
-      // #endif
+#endif
 
       CUDNN_ENFORCE(platform::dynload::cudnnSetTensorNdDescriptor(
           data_desc_, CudnnDataType<T>::type,