TF_Status* status) const {
  std::vector<PartialTensorShape> expected_output_shapes(expected_max_outputs);
  StartExecute(context, inputs, operation_name, attributes,
               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(

  // run and have results. If is_async=true it means that all of the
  // device-specific executors have scheduled the op.
  //
  // Accepts inferred shapes for outputs (`expected_output_shapes`), which if
  // fully defined will avoid querying the shapes of the underlying
  // TensorHandles when ParallelTensor::Shape is called. This allows async
  // computation to continue without blocking.
  //

                               /*expected_max_outputs=*/1,
                               cancellation_manager);
  auto outputs = parallel_device.Join(
      /*expected_output_shapes=*/{PartialTensorShape({})}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;