return trackable_object_graph_;
  }

  /// BundleReader for accessing the variables bundle.
  BundleReader* variable_reader() { return variable_reader_.get(); }

  /// The GraphDebugInfo (or nullptr if none).
  GraphDebugInfo* debug_info() { return debug_info_.get(); }

  /// Restores objects, invoking the callback with the node id in the
  /// saved_object_graph() and the corresponding TrackableObject from the

    const GraphDef& graphdef, const GraphDebugInfo& debug_info,
    const GraphImportConfig& specs, mlir::MLIRContext* context);

// Given a Graph, returns a MLIR module containing the graph, expressed with
// tf_executor dialect.
absl::StatusOr<mlir::OwningOpRef<mlir::ModuleOp>> ConvertGraphToMlir(
    const Graph& graph, const GraphDebugInfo& debug_info,

                      internal::FileExists(Env::Default(), debug_info_pb_path));
  if (debug_info_pb_exists) {
    GraphDebugInfo debug_info;
    TF_RETURN_IF_ERROR(
        ReadBinaryProto(Env::Default(), debug_info_pb_path, &debug_info));
    *debug_info_proto = std::make_unique<GraphDebugInfo>(std::move(debug_info));
  }
  return absl::OkStatus();
}

// it fails to convert the input.
absl::Status ConvertGraphDefToTFLiteFlatBuffer(
    const toco::ModelFlags& model_flags, toco::TocoFlags& toco_flags,
    const GraphDebugInfo& debug_info, const GraphDef& input,
    std::string* result);

}  // namespace tensorflow

                                      MetaGraphDef* meta_graph_def);

// Store debug info from the SavedModel export dir.
Status ReadSavedModelDebugInfoIfPresent(
    absl::string_view export_dir,
    std::unique_ptr<GraphDebugInfo>* debug_info_proto);

}  // namespace tensorflow

inline constexpr char kSavedModelFilenamePbTxt[] = "saved_model.pbtxt";

// Subdirectory where debugging related files are written.
inline constexpr char kSavedModelDebugDirectory[] = "debug";

// File name for GraphDebugInfo protocol buffer which corresponds to the
// SavedModel.
inline constexpr char kSavedModelDebugInfoFilenamePb[] =
    "saved_model_debug_info.pb";

// Directory in which to save the SavedModel variables.

// flag is set to true returns a dictionary that contains string representation
// of the converted model and some statistics like arithmetic ops count.
// `debug_info_str` contains the `GraphDebugInfo` proto. When
// `enable_mlir_converter` is True, use MLIR-based conversion instead of
// TOCO conversion.
PyObject* Convert(PyObject* model_flags_proto_txt_raw,
                  PyObject* toco_flags_proto_txt_raw,

static std::string ImportGraphDefImpl(const std::string& proto,
                                      const std::string& pass_pipeline,
                                      bool show_debug_info,
                                      GraphDebugInfo& debug_info,
                                      GraphImportConfig& specs,
                                      TF_Status* status) {
  GraphDef graphdef;
  auto s = tensorflow::LoadProtoFromBuffer(proto, &graphdef);

  }
  auto convert = [&]() -> Status {
    mlir::StatusScopedDiagnosticHandler status_handler(&context);
    // TODO(jpienaar): Both the graph debug info and import config should be
    // specifiable.
    GraphDebugInfo debug_info;
    switch (config.dialect) {
      case MlirDumpConfig::Dialect::kTFG: {
        TF_ASSIGN_OR_RETURN(module,
                            mlir::tfg::ImportGraphAndFunctionsToMlir(

  EXPECT_FALSE(st.ok());
}

TEST_F(ReaderTest, ReadSavedModelDebugInfoIfPresent) {
  const string export_dir = GetDataDependencyFilepath(TestDataSharded());
  std::unique_ptr<GraphDebugInfo> debug_info_proto;
  TF_ASSERT_OK(ReadSavedModelDebugInfoIfPresent(export_dir, &debug_info_proto));
}

TEST_F(ReaderTest, MetricsNotUpdatedFailedRead) {
  MetaGraphDef meta_graph_def;