How to improve the slow transfer rate of tensors from GPU to CPU

C++
1

How to improve the slow transfer rate of tensors from GPU to CPU in libtorch?

2

How did you profile this operation and what is the time spent for which tensor shape?