std::vector<Output> reserve_spaces;
  reserve_spaces.push_back(op.output(3));
  reserve_spaces.push_back(op.output(4));
  if (op.num_outputs() > 5) {
    reserve_spaces.push_back(op.output(5));
  }

  if (is_training) {
    return grad_fn(scope, x, grad_y, scale, reserve_spaces, epsilon,
                   data_format, is_training, grad_outputs);
  } else {

  %cst_0 = "tf.Const"() {value = dense<0.500000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

     // not matter what value we provide here for replacement.
     /*batch_mean=*/(replaceWithValue $x),
     /*batch_variance=*/(replaceWithValue $x),
     /*reserve_space_1=*/(replaceWithValue $x),
     /*reserve_space_2=*/(replaceWithValue $x),
     /*reserve_space_3=*/(replaceWithValue $x)],
    [(HasNoUseOf:$root__1), (HasNoUseOf:$root__2),
     (HasNoUseOf:$root__3), (HasNoUseOf:$root__4),
     (HasNoUseOf:$root__5)]>;

//    // not matter what value we provide here for replacement.
//      /*batch_mean=*/(replaceWithValue $x),
//      /*batch_variance=*/(replaceWithValue $x),
//      /*reserve_space_1=*/(replaceWithValue $x),
//      /*reserve_space_2=*/(replaceWithValue $x),
//      /*reserve_space_3=*/(replaceWithValue $x)],
//     [(HasNoUseOf:$root__1), (HasNoUseOf:$root__2),
//      (HasNoUseOf:$root__3), (HasNoUseOf:$root__4),

                                /*reserve_space_1=*/reserve_space_1,
                                /*reserve_space_2=*/batch_variance});
      } else {  // TF::FusedBatchNormV3Op
        // For FusedBatchNormV3Op, also create a constant tensor to forward to
        // last reserve_space_3 output.
        auto reserve_space_3_type =

    TensorOf<[TF_Bfloat16, TF_Float16, TF_Float32]>:$y,
    TF_Float32Tensor:$batch_mean,
    TF_Float32Tensor:$batch_variance,
    TF_Float32Tensor:$reserve_space_1,
    TF_Float32Tensor:$reserve_space_2,
    TF_Float32Tensor:$reserve_space_3
  );

  TF_DerivedOperandSizeAttr num_side_inputs = TF_DerivedOperandSizeAttr<5>;
  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;

                                PatternRewriter &rewriter) const override {
    auto new_result_types =
        llvm::to_vector<6>(tf_fused_batch_norm_op.getResultTypes());
    // reserve_space_3
    new_result_types.push_back(
        UnrankedTensorType::get(FloatType::getF32(rewriter.getContext())));
    auto tf_fused_batch_norm_op_v3 = CreateTfOp<TF::FusedBatchNormV3Op>(