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());

// A simple logging device to test custom device registration.
#include <memory>

#include "absl/strings/match.h"
#include "tensorflow/c/c_api.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/c/eager/c_api_test_util.h"
#include "tensorflow/c/eager/custom_device_testutil.h"
#include "tensorflow/c/tf_status.h"

  std::vector<std::string> devices{
      "/job:localhost/replica:0/task:0/device:CPU:0",
      "/job:localhost/replica:0/task:0/device:CPU:1"};
  ParallelDevice parallel_device(std::move(devices));
  {
    std::unique_ptr<ParallelTensor> float_tensors =
        parallel_device.ScalarsFromSequence<float>({10.0, 11.0}, context.get(),

  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 =

  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  const char* first_device_name =
      "/job:localhost/replica:0/task:0/device:CPU:0";
  const char* second_device_name =
      "/job:localhost/replica:0/task:0/device:CPU:1";
  std::array<const char*, 2> underlying_devices{first_device_name,
                                                second_device_name};

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(

  TFE_CustomDevice custom_device;
  custom_device.copy_tensor_to_device = &CopyToLoggingDevice;
  custom_device.copy_tensor_from_device = &CopyTensorFromLoggingDevice;
  custom_device.delete_device = &DeleteLoggingDevice;
  custom_device.execute = &LoggingDeviceExecute;
  LoggingDevice* device = new LoggingDevice;
  device->arrived_flag = arrived_flag;
  device->executed_flag = executed_flag;

// Create and modify a variable placed on a parallel device which composes
// `first_device` and `second_device`.
void BasicTestsForTwoDevices(TFE_Context* context, const char* first_device,
                             const char* second_device) {
  // Register the custom device
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);

  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);

}

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);