}

  auto stateful_string = mlir::TF::TensorFlowDialect::GetStatefulAttrName();
  if (auto attr = function->getAttrOfType<mlir::UnitAttr>(stateful_string)) {
    func_def.mutable_signature()->set_is_stateful(true);
  }

  // Ignore the gradient and is_stateful attribute on the function as they have
  // been handled above. Ignore the entry func attribute as it is an MLIR

  }

  auto stateful_string = mlir::TF::TensorFlowDialect::GetStatefulAttrName();
  if (auto attr = function->getAttrOfType<mlir::UnitAttr>(stateful_string)) {
    func_def.mutable_signature()->set_is_stateful(true);
  }

  // Ignore the gradient and is_stateful attribute on the function as they have
  // been handled above. Ignore the entry func attribute as it is an MLIR

        {"shape_shape:output:0", "minval:output:0", "maxval:output:0"},
        {{"Tout", DataType::DT_INT32}, {"T", DataType::DT_INT32}}}},
      /*ret_def=*/{{"out", "shape:output:0"}});

  func.mutable_signature()->set_is_stateful(true);
  *flib_def.add_function() = std::move(func);
  TF_ASSERT_OK(root.graph()->AddFunctionLibrary(flib_def));
  NodeDef call_node;
  call_node.set_name("fn_call");
  call_node.set_op("Stateful_func");

    auto updated_fdef = std::make_unique<FunctionDef>();
    TF_RETURN_IF_ERROR(
        GraphToFunctionDef(*g, new_func_name, updated_fdef.get()));
    updated_fdef->mutable_signature()->set_is_stateful(true);
    const FunctionDef* original_fdef = fld->Find(func_name);
    if (original_fdef) {
      for (const auto& attr : original_fdef->attr()) {
        (*updated_fdef->mutable_attr())[attr.first] = attr.second;

    .Attr("Tresults: list(type) >= 0")
    .Attr("function: func")
    // XLA random-number generation ops are stateful.
    // TODO(phawkins): create stateful and non-stateful variants of XlaLaunch.
    .SetIsStateful()
    .Doc("XLA Launch Op. For use by the XLA JIT only.");

REGISTER_OP("XlaLaunchV2")
    .Input("args: Targs")
    .Output("results: Tresults")
    .Attr("Targs: list(type) >= 0")

)doc")
    .Attr("container: string = ''")
    .Attr("shared_name: string = ''")
    .Attr("debug_name: string = ''")
    .Attr("dtype: type")
    .Attr("shape: shape")
    .Output("resource: resource")
    .SetIsStateful()
    .SetShapeFn([](shape_inference::InferenceContext* c) {
      c->set_output(0, c->Scalar());
      DataType t;
      TF_RETURN_IF_ERROR(c->GetAttr("dtype", &t));
      PartialTensorShape p;

    .Input("tensor_data: T")
    .Attr("log_dir_path: string")
    .Attr("file_name: string")
    .Attr("T: type")
    .Attr("enabled: bool")
    .Attr("func_name: string")
    .Attr("node_name: string")
    .SetIsStateful();

class DumpTensorOp : public OpKernel {
 public:
  explicit DumpTensorOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
    string log_dir_path;
    string file_name;
    string func_name;

REGISTER_OP("CalibrationStatisticsSaver")
    .Input("args: Tin")
    .Attr("Tin: list(type) >= 0")
    .Attr("ids: list(string) >= 1")
    .Attr("calibration_methods: list(int) >= 1")
    .Attr("output_file_path: string")
    .SetIsStateful()
    .Doc(R"doc(
Aggregates and saves the calibration statistics data.

This op collects outputs of multiples CustomAggregator ops, which includes
`min`, `max` and `histogram`. Then it aggregates them according to the

    reinterpret_cast<OpDefBuilder*>(builder)->func_name();                \
  }

DEFINE_BUILDER_BOOL_SETTER(SetIsCommutative)
DEFINE_BUILDER_BOOL_SETTER(SetIsAggregate)
DEFINE_BUILDER_BOOL_SETTER(SetIsStateful)
DEFINE_BUILDER_BOOL_SETTER(SetAllowsUninitializedInput)

void TF_OpDefinitionBuilderAddAttr(TF_OpDefinitionBuilder* builder,
                                   const char* attr_spec) {