I’ve been working on several PRs that fix serious CUDA compile-time issues (see e.g. #1573 for missing includes, generated headers, test wiring, etc.). While doing that, it became clear that this kind of breakage can go unnoticed because CI does not currently exercise the CUDA build path.
While full runtime CUDA testing is difficult in standard CI (no NVIDIA GPUs available on most public runners), a minimal compile-only check is perfectly feasible and would catch many of these problems early.
I’m not sure how official or optional CUDA support is considered upstream, so I’d rather leave that decision to the maintainers.
If you think it would be useful, I’d be happy to open a PR adding a compile-only CUDA check once the related compile fixes are in place. If not, feel free to close this.
Thanks!
I’ve been working on several PRs that fix serious CUDA compile-time issues (see e.g. #1573 for missing includes, generated headers, test wiring, etc.). While doing that, it became clear that this kind of breakage can go unnoticed because CI does not currently exercise the CUDA build path.
While full runtime CUDA testing is difficult in standard CI (no NVIDIA GPUs available on most public runners), a minimal compile-only check is perfectly feasible and would catch many of these problems early.
I’m not sure how official or optional CUDA support is considered upstream, so I’d rather leave that decision to the maintainers.
If you think it would be useful, I’d be happy to open a PR adding a compile-only CUDA check once the related compile fixes are in place. If not, feel free to close this.
Thanks!