TF_ASSERT_OK_AND_ASSIGN(auto stream, executor->CreateStream());

  se::DeviceMemoryBase gpu_dst{device_tensor.data(), 4 * sizeof(float)};
  xla::Shape shape;
  TF_ASSERT_OK(TensorShapeToXLAShape(DT_FLOAT, TensorShape({2, 2}), &shape));

  // Copy the cpu_tensor to the GPU first before trying to copy it back.
  TF_ASSERT_OK(
      stream->Memcpy(&gpu_dst, origin_cpu_tensor.data(), gpu_dst.size()));