StatusPtr first_bad_status(nullptr);
  for (int device_index = 0; device_index < underlying_devices_.size();
       ++device_index) {
    DeviceThread* device_thread = device_threads_[device_index].get();
    per_device_output_tensors.push_back(device_thread->Join(status));
    // We will run every Join even if there are bad statuses in case the user
    // wants to recover and continue running ops on the parallel device (which

  // TODO(allenl): Keep a map from outer thread to list of inner threads rather
  // than a single list of threads so aliased nested parallel devices don't
  // re-use a thread.
  std::vector<std::unique_ptr<DeviceThread>> device_threads_;
  // A cancellation manager to use if the caller does not provide one. When ops
  // are executed asynchronously this must outlive the queued op, so it can't be
  // function-local to Execute.