What do you mean you were even able to make it work? It works out of the box with no configuration.
Can't speak for CUDA specifically, because I don't do anything with LLMs or ML locally, but I did have to rebuild ffmpeg with some some bits built with nvcc (and maybe an nvidia provided library- sorry it's been a while) to get GPU accelerated transcoding going.
But yes, the GPU passthrough layer in WSL2 (or however it's actually implemented) is present by default.
I meant compared to other Linux/Windows coexistence tools I've used in the past where GPU passthrough was a huge challenge
Yeah, except that requires the entire hyperv and constantly running into memory issues.
Wsl1 on the other hand is way too slow on some use case.
I stay with cygwin for common shell, wsl1 for simple linux user space tools, and ssh to a remote system if I need a real Linux.
WSL2 has options now to reclaim unused memory and it has never required a full HyperV just a limited subset of it. That's how and why you can use it on Windows Home.