I just got DirectML working with torch in WSL2 which was fairly painless.

I am wondering if that isn’t a better option than trying to emulate CUDA directly. Love it or hate it, Microsoft did do a fairly good job wrangling in different platforms with DirectX.