writing an ebuild for llama.cpp, tests with cuda fail

[Arniiiii]

Hello.

This started from a discord guy asking for a ebuild for llama.cpp at gentoo's discord : https://discord.gg/gentoolinux https://discord.com/channels/249111029668249601/1335943645052014602

I'm trying to write an ebuild for llama.cpp ( https://github.com/ggerganov/llama.cpp ) :
https://github.com/Arniiiii/ex_repo/blob/master/dev-cpp/llama-cpp/llama-cpp-0_pre4676.ebuild .

There's a problem:

if `USE=cuda`, tests fail with error like "failed to initialize CUDA: OS call failed or operation not supported on this OS" : test log: https://paste.gentoo.zip/cQhxxPaF .

But a first failed test passes just fine when run as root : https://paste.gentoo.zip/mz3vZ4Vo

Full build.log : https://paste.gentoo.zip/n4VbzZw2
emerge --info : https://paste.gentoo.zip/WTPVC1mz

I've tried searching for ebuild that use `cuda` eclass and took some lines from pycuda from src_test() , but it didn't help.
I've tried searching the error at bugs.gentoo.org , the forum and just in the Internet , however I didn't found something related/helpful

Looks like sandbox / permission issue.

I use my ebuild in the way to test: