* Since GPU is an async device by default. We should sync computation
  when Python invoke `run`. So Python can get the correct computation
  result