// `send_device`, `recv_device`, and `send_device_incarnation` set.
TFE_Op* SendOp(TFE_Context* ctx, TFE_TensorHandle* in,
               const std::string& op_name, const std::string& send_device,
               const std::string& recv_device,
               tensorflow::uint64 send_device_incarnation);

// Return a RecvOp op `op_name` with the attributes `send_device`,

  TF_DeleteStatus(status);

  TFE_OpSetAttrType(op, "T", TFE_TensorHandleDataType(in));
  TFE_OpSetAttrString(op, "tensor_name", "dummy", 5);
  TFE_OpSetAttrString(op, "send_device", send_device.c_str(),
                      send_device.size());
  TFE_OpSetAttrString(op, "recv_device", recv_device.c_str(),
                      recv_device.size());
  TFE_OpSetAttrInt(op, "send_device_incarnation", send_device_incarnation);

      second_components[0].get(), status.get());
  ASSERT_EQ(second_underlying_devices[0], first_device);
  std::string second_device = TFE_TensorHandleBackingDeviceName(
      second_components[1].get(), status.get());
  ASSERT_EQ(second_underlying_devices[1], second_device);

  // Un-pack the first parallel device's tensor too
  std::array<TensorHandlePtr, 2> first_components;

  const char* first_device = "/job:worker/replica:0/task:1/device:CPU:0";
  const char* second_device = "/job:worker/replica:0/task:2/device:CPU:0";
  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> underlying_devices{first_device, second_device};
  RegisterParallelDevice(context.get(), device_name, underlying_devices,
                         status.get());