Libtorch 100% CPU load

I am trying to use libtorch to do some inference with CUDA.
I have noticed that CPU load is almost 100% all the time. I have Intel Core i7-9700K.
When I run Visual Studio Profiler it shows that functions like torch::stack load CPU massively.

Currently I am using official libtorch Release build.
I suppose that probably I should built libtorch from sources with other options.

Am I correct? And if yes, which options should I enable?