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());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  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<TFE_Op, decltype(&TFE_DeleteOp)> handle_op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);

}

TF_Buffer* TF_CreateConfig(unsigned char enable_xla_compilation,
                           unsigned char gpu_memory_allow_growth,
                           unsigned int num_cpu_devices) {
  tensorflow::ConfigProto config;
  auto* optimizer_options =
      config.mutable_graph_options()->mutable_optimizer_options();
  if (enable_xla_compilation) {

  TF_DeviceList* devices = TFE_ContextListDevices(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_DeleteContext(ctx);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  const int num_devices = TF_DeviceListCount(devices);
  EXPECT_GE(num_devices, 1) << "At least one CPU device should exist";
  for (int i = 0; i < num_devices; ++i) {

  float* m_float = static_cast<float*>(TF_TensorData(m_data));
  TF_DeviceList* devices = TFE_ContextListDevices(ctx, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  int num_devices = TF_DeviceListCount(devices);
  for (int d = 0; d < num_devices; ++d) {
    const char* name = TF_DeviceListName(devices, d, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TF_DeviceList* device_list = list.get();

  CHECK_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  const int num_devices = TF_DeviceListCount(device_list);
  LOG(INFO) << "There are " << num_devices << " devices.";
  for (int i = 0; i < num_devices; ++i) {
    const char* device_name = TF_DeviceListName(device_list, i, s);
    CHECK_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  ASSERT_TRUE(executed);
  TFE_DeleteTensorHandle(retval);

  // Custom device: inputs in different custom devices fails.
  matmul.reset(MatMulOp(context.get(), hcustom0.get(), hcustom1.get()));
  num_retvals = 1;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  ASSERT_NE(TF_OK, TF_GetCode(status.get()));

      TFE_ContextListDevices(context.get(), status.get()), TF_DeleteDeviceList);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  bool has_tpu = false;
  for (int device_index = 0; device_index < TF_DeviceListCount(devices.get());
       ++device_index) {
    std::string device_type =
        TF_DeviceListType(devices.get(), device_index, status.get());

  }
}

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(

  std::vector<std::string> underlying_devices_vector;
  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 =