Zsgpu

[joshkamm]

Paul plans to work on this branch for some time and use it as the dependency for his replacement of FancyElectrons. It is being kept separate because Paul's changes likely will not continue to be compatible with FancyElectrons. Once the replacement is complete, the plan is to merge this pr and pin FancyElectrons to a commit before this pr was merged. Ultimately, Paul plans to migrate all FancyElectrons usage in the group to his replacement.

I'm creating this pr well before it's ready to merge as a space for notes and to monitor progress.

• Merge main and zsgpu into a temporary branch
• Comment out all occurrences of async
• Create a separate branch to pin FancyElectrons to before merging this branch (add note in FancyElectrons readme)
• Update the gh actions ci to use one of the newer examples (update in pixi, gh actions script, and maybe cmake)
• Reapply SSTRF changes
• Update version number

For Alex and Nate: do the examples currently in the repository look representative of what would be helpful for validating the accuracy of the code?

Vaibhav's testing

• RI = 3 doesn't work
• Testing RI = 4

Closes #16
Closes #17

[paulzim46]

It would be nice to fix the multiGPU operation with the latest nvidia compilers, but I also suspect we just have a compiler bug on our hands. Versions 24.9 and 24.11 do not work; 20.7 is fine.

[joshkamm]

Hmm @paulzim46 do you think it's worth trying to sufficiently isolate it to provide the compiler team with a lead if it is a bug?

While working on the infrastructure, nvhpc has also been the most difficult dependency to manage. It seems like their licensing makes it difficult for a package manager to redistribute.

I've noticed that gcc also has openacc support which may be easier to manage from an infrastructure angle, but I don't know whether it's as reliable and performant in general, or how many subtle differences there are between them.