macOS GitHub Actions CI stuck on adding AMDGPU.jl

[williamfgc]

Mostly asking for guidance and to know if this is a known issue.

Our macOS CI on GitHub Actions runners shows a spurious behavior getting stuck on AMDGPU.jl fairly frequently see this log and other recent failure log.

Is there anything that could explain this behavior on the packaging side? Thanks in advance!

[luraess]

Why would you add AMDGPU on macOS? Can't you run AMDGPU backend tests only on Linux targets? Sorry if I am missing something.

[williamfgc]

@luraess I just went with the assumption that GPU back ends are lightweight packages (not different from adding CUDA.jl to a non-NVIDIA GPU system from their docs on Conditional Use). Because CUDA.jl always loads, even if the user doesn't have a GPU or CUDA, you should just depend on it like any other package (and not use, e.g., Requires.jl). The strange thing is that this works sometimes and sometimes doesn't with AMDGPU.jl, it always works with CUDA.jl on our CI. I just wanted to know if this is known.

Weak dependencies don't propagate to tests and I must admit I don't know of a "best practice" to put these two together and I'd glad to listen to options.

From the docs here it seems a separate test/Project.toml still TBD:

The exact interaction between Project.toml, test/Project.toml and their corresponding Manifest.tomls are not fully worked out and may be subject to change in future versions. The older method of adding test-specific dependencies, described in the previous section, will therefore be supported throughout all Julia 1.X releases.

Also, this discussion might be related.

[pxl-th]

Closing this as we now run our CI on GPU-less machines and it passes:
https://buildkite.com/julialang/amdgpu-dot-jl/builds/2911#019556bf-1b5d-48be-9864-fed4f121f20f

Feel free to re-open if you experience this issue again.