TF_Message(first_bad_status.get()));
    return result;
  }
  // For each output of the original operation, pack the per-device
  // TensorHandles we've computed into a single parallel TensorHandle.
  std::vector<std::unique_ptr<ParallelTensor>> per_device_outputs;
  per_device_outputs.reserve(first_op_output_count);
  for (int i = 0; i < first_op_output_count; ++i) {

    TFE_DeleteTensorHandle(n);
  }

  int num_retvals = 1;
  if (async) {
    // Enqueue dummy ops so we backlog async execution & actually test async.
    // This is usually unnecessary, but we've experienced the occasional test
    // failure when testing async mode with no explicit forwarding.
    for (int i = 0; i < 100000; ++i) {
      TFE_Op* add_op_dummy = AddOp(ctx, m, m);