Support for OpenMPI 5.x: `--npernode` flag removed in `kv_cache` and `vectordb` workloads

[wangjiale01]

Hi team,

I am trying to run the kv_cache and vectordb workloads in an environment with OpenMPI 5.0.10. The execution fails because the launch scripts still use the --npernode flag, which was officially removed in OpenMPI 5.0.0.

My question is:
Is there any plan to update the scripts to support OpenMPI 5.x (e.g., replacing --npernode with --map-by ppr:<N>:node) in the near future?

I am asking to decide whether I should wait for the official update or just downgrade my OpenMPI to 4.x to proceed with my testing right now.

Thanks!

[dslik]

I would recommend proceeding with the downgrade to OpenMIP 4.x option at this time.

We'll want to keep this open, though, so it can get fixed after this round is over.

[FileSystemGuy]

@wangjiale01 Yes, its too late in the submission cycle to take on a new major version of OpenMPI. Was there a particular reason you could not use the OpenMPI version that is specified in the uv.lock file?

[wangjiale01]

Yes, I've encountered some unexplained issues with Open MPI 4.x in my environment. Would it be okay if I use version 5.0.10 to test other use cases for my submission?