Richard Berger

@bernhardkaindl FWIW, I've already tested this on multiple Cray systems it's working well for my use cases. Probably the only thing missing is the OneAPI case.

@etiennemlb I made the proposed changes here: https://github.com/rbberger/spack/tree/feature/cray-mpich-gtl

Closing, since this should be a non-issue with a corrected configuration as mentioned above. @dipietrantonio reopen if you disagree.

Another data point. Based on this https://developer.nvidia.com/blog/using-cuda-stream-ordered-memory-allocator-part-2/#interprocess_communication_support it seems that `cudaIpcGetMemHandle` (which `mpich`, `cray-mpich`, `openmpi`, `libfabric` use interally) won't work with `cudaMallocAsync` memory at all. Apparently only UCX is aware...

I'm noticing a strange new error during mirror creation that happens on `develop` when I have `git` as one of the root specs in an environment and define it as...

> C++17 compliant compiler @lindsayad NVCC is not such a compiler. Kokkos uses that by default when configuring it for CUDA. However, since we need a real C++ compiler, you...

@lammps/core FYI

@akohlmey FYI. Note that I needed to add installation steps in the `examples/plugins/CMakeLists.txt`. I made the choice to install to `/lib{64}/lammps/plugins`. Any plugin that follows that convention will be automatically...

@bernhardkaindl thanks for testing this. Since the last update there have been more changes to these packages we use internally. I'll incorporate those and fixes for your findings in another...

Closing this one. Will recreate it in spack-packages, but one PR for each new package.