expected_max_outputs, *default_cancellation_manager_);
  auto result = Join(expected_output_shapes, status);
  if (TF_GetCode(status) != TF_OK) {
    std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> await_status(
        TF_NewStatus(), TF_DeleteStatus);
    // Wait until all pending nodes have completed since they may have a
    // reference to default_cancellation_manager_. We ignore the status return

  // are executed asynchronously this must outlive the queued op, so it can't be
  // function-local to Execute.
  mutable std::unique_ptr<CancellationManager> default_cancellation_manager_;
};

// Contains a tuple of tensors, one on each of the `underlying_devices_` of the
// ParallelDevice.
class ParallelTensor {
 public: