Open ConnorBaker opened 8 months ago
Why would these packages take a cudaSupport argument if they only exist as cuda variants?
IIRC there are some packages that try using CUDA at runtime and fall back on CPU. Admittedly I don't remember who's doing this (maybe blender or something like that?), or if they're source builds.
But I'm not opposed to marking a source-built package broken if there's no reasonable way the build would succeed without cudaSupport
. In that case I guess @SomeoneSerge 's question stands: "why accept a cudaSupport
argument if they have no other reasonable configuration?"
IIRC there are some packages that try using CUDA at runtime and fall back on CPU. Admittedly I don't remember who's doing this (maybe blender or something like that?), or if they're source builds.
Not in the cudaPackages
namespace though?
Oh true! None in the cudaPackages namespace thankfully
For source-available packages like CUDA samples, NCCL, NCCL-Tests, and Saxpy, we should mark them as broken if
cudaSupport
is false. My reasoning is this: when source-based packages generate code for GPUs, we should make certaincudaSupport
is enabled.However, binary-based packages should not be marked as broken if
cudaSupport
is false (not explicitly anyway -- it's fine if they're broken due to a dependency on a package marked as broken) because they do not generate (or perhaps even contain!) GPU device code. As an example,cuobjdump
doesn't require a GPU to run and should be usable even ifcudaSupport
is false (assumingallowUnfree
is true).@NixOS/cuda-maintainers @SomeoneSerge @samuela thoughts?
Add a :+1: reaction to issues you find important.