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());

  }

  auto* gpu_options = config.mutable_gpu_options();
  gpu_options->set_allow_growth(gpu_memory_allow_growth);

  (*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the

ParallelDevice::ParallelDevice(const std::vector<std::string>& devices,
                               bool is_async, int in_flight_nodes_limit)
    : underlying_devices_(devices),
      default_cancellation_manager_(absl::make_unique<CancellationManager>()) {
  device_threads_.reserve(devices.size());
  for (int device_index = 0; device_index < devices.size(); ++device_index) {
    device_threads_.emplace_back(new DeviceThread(

  underlying_devices_vector.reserve(num_underlying_devices);
  for (int device_index = 0; device_index < num_underlying_devices;
       ++device_index) {
    underlying_devices_vector.push_back(underlying_devices[device_index]);
  }
  std::unique_ptr<ParallelDevice> parallel_device(
      new ParallelDevice(underlying_devices_vector));
  *device_info =

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  const DLTensor* dltensor_out = &dlm_out->dl_tensor;
  EXPECT_EQ(dltensor_out->device.device_type, dltensor_in->device.device_type);
  EXPECT_EQ(dltensor_out->device.device_id, dltensor_in->device.device_id);
  EXPECT_EQ(dltensor_out->ndim, dltensor_in->ndim);
  EXPECT_EQ(dltensor_out->dtype.code, dltensor_in->dtype.code);

  // Remote device on `worker1`.
  const char remote_device[] = "/job:localhost/replica:0/task:1/device:CPU:0";
  // `ctx_0`, `ctx_1`, `ctx_2` contains `remote_device`.
  {
    const std::vector<std::string>& device_names = ListDeviceNames(ctx_0);
    ASSERT_TRUE(std::find(device_names.begin(), device_names.end(),
                          remote_device) != device_names.end());
  }

  {

}

void TF_SetRequestedDevice(TF_Graph* graph, TF_Operation* op,
                           const char* device) {
  using tensorflow::RecordMutation;
  mutex_lock l(graph->mu);
  op->node.set_requested_device(device);
  RecordMutation(graph, *op, "setting device");
}

void TF_RemoveAllControlInputs(TF_Graph* graph, TF_Operation* op) {
  mutex_lock l(graph->mu);

void TF_AbstractOpSetOpType(TF_AbstractOp* op, const char* const op_type,
                            TF_Status* s) {
  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Reset(op_type,
                           /*raw_device_name=*/nullptr));
}

void TF_AbstractOpSetOpName(TF_AbstractOp* op, const char* const op_name,
                            TF_Status* s) {
  TracingOperation* tracing_op = dyn_cast<TracingOperation>(unwrap(op));

                      send_device.size());
  TFE_OpSetAttrString(op, "recv_device", recv_device.c_str(),
                      recv_device.size());
  TFE_OpSetAttrInt(op, "send_device_incarnation", send_device_incarnation);

  return op;
}

bool GetDeviceName(TFE_Context* ctx, string* device_name,
                   const char* device_type) {

  // Remote device on `worker2`.
  const char remote_device[] = "/job:localhost/replica:0/task:2/device:CPU:0";
  // `ctx_0`, `ctx_1` contains `remote_device`.
  {
    const std::vector<std::string>& device_names = ListDeviceNames(ctx_0);
    ASSERT_TRUE(std::find(device_names.begin(), device_names.end(),
                          remote_device) != device_names.end());
  }

  {