My libtorch build is much slower than prebuilt ones

I am trying to replicate building libtorch with dependencies. I use the instructions on, but still there are some performance degradation. The performance loss is even worse when I try a different GPU other than the one that I used for building the binraries.

I have two questions:

  1. Is the is the official pytorch project that is used for building libtorch? It seems that the instructions are a bit outdated.

  2. Is there any place to see the build log of pre-built libtorch binaries. For instance, which version of cudnn is used, what are the dependencies, etc.