std::vector<TFE_TensorHandle*> handles_task2;
  for (auto* h_task0 : handles_task0) {
    handles_task2.push_back(
        TFE_TensorHandleCopyToDevice(h_task0, ctx, task2_name, status));
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  }

  auto* h1_task2 =
      TFE_TensorHandleCopyToDevice(h1_task0, ctx, task2_name, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

  const char remote_device_name[] =
      "/job:localhost/replica:0/task:1/device:CPU:0";
  auto* h0_task1 =
      TFE_TensorHandleCopyToDevice(h0_task0, ctx, remote_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  auto* h1_task1 =
      TFE_TensorHandleCopyToDevice(h1_task0, ctx, remote_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_TensorHandle* hcpu = TestMatrixTensorHandle(context);
  ASSERT_FALSE(arrived);
  TFE_TensorHandle* hdevice =
      TFE_TensorHandleCopyToDevice(hcpu, context, name, status.get());
  ASSERT_TRUE(arrived);
  ASSERT_FALSE(executed);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  auto* retval_task0 =
      TFE_TensorHandleCopyToDevice(retvals[0], ctx, local_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  CheckTFE_TensorHandleHasFloats(retval_task0, {7, 10, 15, 22});

  TFE_DeleteTensorHandle(retval_task0);

  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);
    TFE_TensorHandle* copy = TFE_TensorHandleCopyToDevice(m, ctx, name, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    void* data = TFE_TensorHandleDevicePointer(copy, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

    // Copy to device
    TFE_TensorHandle* hdevice =
        TFE_TensorHandleCopyToDevice(hcpu, ctx, name.c_str(), status.get());
    if (TF_GetCode(status.get()) != TF_OK) {
      ADD_FAILURE() << tag << " -- " << TF_Message(status.get());
      continue;
    }
    // Copy from device to the same device.
    TFE_TensorHandle* hdevice2 =
        TFE_TensorHandleCopyToDevice(hdevice, ctx, name.c_str(), status.get());

                                      TF_Status* status, void* device_info) {
  LoggingDevice* dev = reinterpret_cast<LoggingDevice*>(device_info);
  TFE_TensorHandle* t = TFE_TensorHandleCopyToDevice(
      tensor, context, dev->underlying_device.c_str(), status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  auto dst = std::make_unique<LoggedTensor>(t);
  *(dev->arrived_flag) = true;

  const char remote_device_name[] =
      "/job:localhost/replica:0/task:1/device:CPU:0";
  auto* h0_task1 =
      TFE_TensorHandleCopyToDevice(h0_task0, ctx, remote_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  auto* h1_task1 =
      TFE_TensorHandleCopyToDevice(h1_task0, ctx, remote_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

          unwrapped_op,
          reinterpret_cast<tensorflow::ImmediateExecutionTensorHandle**>(
              retvals),
          num_retvals);
}

TFE_TensorHandle* TFE_TensorHandleCopyToDevice(TFE_TensorHandle* h,
                                               TFE_Context* ctx,
                                               const char* device_name,