{mpi}[NVHPC/24.1-CUDA-12.4.0] OpenMPI v4.1.6 w/ CUDA 12.3.0

[tanmoy1989]

(created using eb --new-pr)

[Thyre]

Is there any specific reason for CUDAcore?

[tanmoy1989]

No, I guess that was a typo from my side.

[tanmoy1989]

I have addressed your comments above, thanks for checking!

[tanmoy1989]

@Thyre: Any updates on this, please?

[Thyre]

The filename should be OpenMPI-4.1.6-NVHPC-24.1-CUDA-12.4.0.eb. This is why the checks failed for the PR.

---

Aside from that the PR looks fine to me. However, I'm not able to approve or merge this PR. In addition, we should at least do some testing. I hopefully can check on my machines on Friday.

[Thyre]

NVHPC/24.1-CUDA-12.4.0 does not exist, but NVHPC/24.1-CUDA-12.3.0 does. It's probably the easiest solution to switch to CUDA 12.3.0.

[tanmoy1989]

I changed it to CUDA-12.3.0. But UCC-CUDA and UCX-CUDA corresponding versions are available only for CUDA-12.4.0. So that's an issue. Also, I am not sure without the versionsuffix = 'CUDAcore-12.4.0' (which I had on my initial .eb file), how the '-CUDA-%(cudaver)s' command is going to executed for those two dependencies.

[Thyre]

I changed it to CUDA-12.3.0. But UCC-CUDA and UCX-CUDA corresponding versions are available only for CUDA-12.4.0. So that's an issue. Also, I am not sure without the versionsuffix = 'CUDAcore-12.4.0' (which I had on my initial .eb file), how the '-CUDA-%(cudaver)s' command is going to executed for those two dependencies.

Well that's annoying...
If we want to stick to officially supported CUDA versions, we can only choose CUDA 12.3.0 or CUDA 11.8.0. Since NVHPC uses 12.3.0 already, I would stick with that. The only solution I see then is to add additional versions of UCX-CUDA and UCC-CUDA for this CUDA version, which should be easy to do.

[tanmoy1989]

@Thyre But EB already have UCX-CUDA based on CUDA-12.4.0 (link). I am bit confused now! Do I need to write UCX-CUDA-1.15.0 based on CUDA-12.3.0 and same for UCC-CUDA? Same for NCCL (which is a dependency of UCC-CUDA) where EB has it with CUDA-12.4.0 (link).
And I am getting an error while deploying NCCL/CUDA-12.3.0 saying "error: #error -- unsupported GNU version! gcc versions later than 12 are not supported! The nvcc flag '-allow-unsupported-compiler' can be used to override this version check; however, using an unsupported host compiler may cause compilation failure or incorrect run time execution. Use at your own risk.". NCCL/CUDA-12.4.0 got built fine though.