the running_mean and running_var of bn would not be synchronized in dist,
so which leads to bug that eval loss in training is inconsistent with eval only.