ParallelDevice::ParallelDevice(const std::vector<std::string>& devices,
                               bool is_async, int in_flight_nodes_limit)
    : underlying_devices_(devices),
      default_cancellation_manager_(absl::make_unique<CancellationManager>()) {
  device_threads_.reserve(devices.size());
  for (int device_index = 0; device_index < devices.size(); ++device_index) {
    device_threads_.emplace_back(new DeviceThread(

  TF_DataType datatype_enum(
      static_cast<TF_DataType>(DataTypeToEnum<DataType>().value));
  for (int device_index = 0; device_index < num_underlying_devices();
       ++device_index) {
    auto device_value = absl::make_unique<DataType>();
    *device_value = values[device_index];
    std::unique_ptr<TF_Tensor, decltype(&TF_DeleteTensor)> tensor(
        TF_NewTensor(
            datatype_enum, /*dims=*/nullptr, /*num_dims=*/0,

}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(
    const TF_Graph* fn_body, const char* fn_name, int ninputs,
    const TF_Output* inputs, std::vector<OutputTensor>* input_tensors,

      TFE_TensorHandleDevicePointer(tensor, status));
  if (TF_GetCode(status) != TF_OK) return nullptr;
  if (parallel_tensor->num_tensors() == 1) {
    // Copy-off for single-device tensors is allowed to make debugging dynamic
    // control flow easier.
    return TFE_TensorHandleCopySharingTensor(parallel_tensor->tensor(0),
                                             status);
  } else {
    TF_SetStatus(