(*node_def->mutable_attr())["_handle_dtypes"] = handle_dtypes_attr;
      (*node_def->mutable_attr())["_handle_shapes"] = handle_shapes_attr;
    }
  }

  TF_RETURN_IF_ERROR(
      SetShapeAttribute("_output_shapes", arg_type, node_def->mutable_attr()));

  DataType dtype;
  TF_RETURN_IF_ERROR(ConvertToDataType(arg_type.getElementType(), &dtype));
  AttrValue type_attr;
  type_attr.set_type(dtype);

  // the shape inference pass is run early in the pass pipeline, shape inference
  // during import is not necessary.
  config.enable_shape_inference = false;
  // Some graphs may require _output_shapes (an unregistered attribute)
  // to override shapes. It is unfortunately not always set correctly so only
  // do it optionally.
  config.unconditionally_use_set_output_shapes =

func.func @testUnimplementedOp() -> (tensor<i32>, tensor<i32>) {
  %0 = arith.constant dense<1> : tensor<i32>
  %1 = arith.constant dense<2> : tensor<i32>
  %2 = "tf.Maximum"(%0, %1) {_output_shapes = ["tfshape$"]} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  %3 = "tf.Minimum"(%0, %1) {random_attr = "hello"} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  func.return %2, %3: tensor<i32>, tensor<i32>

  // CHECK-DAG: [[LINSPACE:%.*]] = chlo.broadcast_add [[MUL]], [[START]] {broadcast_dimensions = array<i64>}
  // CHECK: return [[LINSPACE]]
  %0 = "tf.Const"() {_output_shapes = ["tfshape$"], device = "", dtype = i32, value = dense<4> : tensor<i32>} : () -> tensor<i32>
  %1 = "tf.LinSpace"(%arg0, %arg1, %0) : (tensor<f32>, tensor<f32>, tensor<i32>) -> tensor<4xf32>
  func.return %1 : tensor<4xf32>

func.func private @gives_a_name_to_anonymous_iterators() {
  // CHECK: "tf.Iterator"
  // CHECK-SAME: output_shapes{{.*}}200x28x28x1{{.*}}200x10
  // CHECK-SAME: output_types = [f32, f32]
  // CHECK-SAME: shared_name = "_iterator1"
  %0 = "tf.AnonymousIteratorV3"() {output_shapes = [
    #tf_type.shape<200x28x28x1>,
    #tf_type.shape<200x10>], output_types = [f32, f32]} : () -> tensor<!tf_type.resource>

    %3 = "tf.RangeDataset"(%0, %1, %2) {device = "/device:CPU:0", output_shapes = [#tf_type.shape<>], output_types = [i64], metadata = ""} : (tensor<i64>, tensor<i64>, tensor<i64>) -> tensor<!tf_type.variant>
    // CHECK: tfrt_fallback_async.executeop key({{[0-9]+}}) cost({{.*}}) device("/device:CPU:0") "tf.FlatMapDataset"({{.*}}) {Targuments = [], metadata = "", output_shapes = [#corert.shape<>], output_types = [i64]} {f = "funcB"} : 1

    %3 = "tf.RangeDataset"(%0, %1, %2) {device = "/device:CPU:0", output_shapes = [#tf_type.shape<>], output_types = [i64], metadata = ""} : (tensor<i64>, tensor<i64>, tensor<i64>) -> tensor<!tf_type.variant>

  %3 = "tf.TensorSliceDataset"(%2) {device = "", output_shapes = [#tf_type.shape<>], metadata = ""} : (tensor<3xi32>) -> tensor<*x!tf_type.variant>
  // CHECK: "tf.MapAndBatchDataset"(%[[TSLICE]], %[[BSIZE:.*]], %[[NPC]]
  // CHECK-SAME: f = @"__inference_Dataset_map_<lambda>_80",
  %4 = "tf.MapDataset"(%3) {device = "",
           f = @"__inference_Dataset_map_<lambda>_80",
           output_shapes = [#tf_type.shape<>], output_types = [i32],

  auto node_index = g.BuildNodeNameIndex();
  Node *add_node = node_index["add"];
  std::vector<PartialTensorShape> output_shapes;
  TF_CHECK_OK(GetNodeAttr(add_node->attrs(), kXlaInferredShapesAttrName,
                          &output_shapes));
  EXPECT_EQ(output_shapes.size(), 1);
  TensorShapeProto shape_proto;
  output_shapes[0].AsProto(&shape_proto);
  EXPECT_EQ(shape_proto.dim_size(), 1);
  EXPECT_EQ(shape_proto.dim(0).size(), 2);
}

  SmallVector<int64_t, 2> output_shape{1, mlir::ShapedType::kDynamic};
  EXPECT_EQ(mlir::cast<RankedTensorType>(output_types[0]).getShape().size(),
            output_shape.size());
  for (int i = 0; i < output_shape.size(); i++) {
    EXPECT_EQ(mlir::cast<RankedTensorType>(output_types[0]).getDimSize(i),
              output_shape[i]);
  }
}