There is some work to be done on the SHTns_jll side. At the moment using one environment for a heterogeneous cluster with pure CPU nodes and GPU nodes leads to recompilation issues and not finding CUDA when it should. Using a local CUDA toolkit also breaks SHTns_jll (see also JuliaPackaging/Yggdrasil#11523 after having done JuliaPackaging/Yggdrasil#9813).
You can compile SHTns yourself, using the local CUDA toolkit. However, this does not solve the issue of heterogeneous architecture. The locally compiled shtns library will need CUDA, or one has to load different compiled libraries depending on CPU-only or GPU.
There is some work to be done on the
SHTns_jllside. At the moment using one environment for a heterogeneous cluster with pure CPU nodes and GPU nodes leads to recompilation issues and not finding CUDA when it should. Using a local CUDA toolkit also breaksSHTns_jll(see also JuliaPackaging/Yggdrasil#11523 after having done JuliaPackaging/Yggdrasil#9813).You can compile SHTns yourself, using the local CUDA toolkit. However, this does not solve the issue of heterogeneous architecture. The locally compiled shtns library will need CUDA, or one has to load different compiled libraries depending on CPU-only or GPU.