std::unique_ptr<ParallelTensor> parallel_tensor(
            parallel_device.CopyToParallelDevice(
                context, absl::get<TFE_TensorHandle*>(input), status));
        if (TF_GetCode(status) != TF_OK) return absl::nullopt;
        parallel_inputs.push_back(parallel_tensor.get());
        implicitly_broadcast_tensors.emplace_back(std::move(parallel_tensor));
      } else {
        TF_SetStatus(