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) {

                              const char* device_name, void* device_info,
                              TF_Status* status) {
  // Fill in default values for optional functionality.
  if (device.pack == nullptr) {
    device.pack = &DefaultCustomDevicePack;
  }
  auto custom_device = std::make_unique<tensorflow::CustomDeviceAPI>(
      ctx, device, device_info, device_name);

  // returned by DeviceName should be "/device:GPU:*" until a particular GPU is
  // chosen for the operation by the device placement logic in the
  // executor. After that, the value returned by DeviceName will be a full
  // device name such as "/job:localhost/replica:0/task:0/device:GPU:1".
  virtual const string& DeviceName() const = 0;

  // Sets the operation device name.
  //

// C API for device. The API is under active development and eventually
// should allow registering a plugin device with TensorFlow.

#ifdef __cplusplus
extern "C" {
#endif

// TF_Device is a C wrapper to the C++ TF Device class. This is to be passed
// through TF_OpKernelContext, and is opaque to plugin.
typedef struct TF_Device TF_Device;

typedef struct TF_VariableInfo TF_VariableInfo;

                                      DEVICE_CPU))));
}

TEST(TensorPjRtBufferUtilTest, GetPjRtCApiClientIncorrectType) {
  TF_ASSERT_OK_AND_ASSIGN(
      auto pjrt_client,
      xla::GetTfrtCpuClient(/*asynchronous=*/true, /*cpu_device_count=*/1));
  TF_ASSERT_OK(SetPjRtClientInTFGlobalResourceManager(DEVICE_CPU,
                                                      std::move(pjrt_client)));

  // 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());
  }

  {

  // Check that copying it to the old remote device (named localhost) fails.
  TFE_TensorHandleCopyToDevice(h0_task0_new, ctx, remote_device_name, status);
  EXPECT_NE(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Copying and executing on the new remote device works.
  const char new_remote_device_name[] =
      "/job:worker/replica:0/task:1/device:CPU:0";
  const char new_local_device_name[] =

  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);

    device_id = parsed_name.id;
  }

  ctx.device_id = device_id;
  if (device_type == "CPU") {
    ctx.device_type = DLDeviceType::kDLCPU;
  } else if (device_type == "GPU") {
#if TENSORFLOW_USE_ROCM
    ctx.device_type = DLDeviceType::kDLROCM;
#else
    ctx.device_type = DLDeviceType::kDLCUDA;
#endif
  } else {
    status->status = tensorflow::errors::InvalidArgument(

}

func runDevice(appdir, bundleID string, args []string) error {
	attempt := 0
	for {
		// The device app path reported by the device might be stale, so retry
		// the lookup of the device path along with the lldb launching below.
		deviceapp, err := findDeviceAppPath(bundleID)
		if err != nil {
			// The device app path might not yet exist for a newly installed app.
			if attempt == 5 {
				return err
			}
			attempt++