Feature Request: Force-Off ROCm and CUDA builds in gen_linux.sh even if they are present.

[dreirund]

Ahoj,

there is a need to force a build without CUDA or ROCm, even if some of their libraries are present on the system.

But current gen_linux.sh forces ROCm build if some libraries are found, even if ROCm is not desired.

Arch Linux AUR package ollama-nogpu-git has a patch to patch out all CUDA and ROCm stuff in gen_linux.sh, but this upstream here seems to change things too fast that it is not feasible to stay up to date with that patch.

So I here wish for a feature to make it built-in that ROCm (and CUDA) build can be forced-off (e.g. by an environment variable or command line switch) even if their libraries or just some of them are present. (On my system, I have only some stuff in /opt/rocm/, which is needed by some package, but is not a full and not a functional ROCm installation. This currently breaks building ollama.)

Regards!

[dhiltgen]

This was recently fixed via #4462 to use the same variables we have on the windows build.

[dreirund]

I see you have added ONEAPI build. Can you please also add a force kill switch to disable ONEAPI build even if `/opt/intel/oneapi` is present? (It is, but I do not have the "icx"/ "icpx" compiler and won't install it because of big size.) And actually for _each_ new build target you add that gets auto detected add a kill switch to forcefully _not_ build it? Regards!

[dhiltgen]

@dreirund good catch! We missed that on the recent addition of oneapi - I'll get a PR up to add that.