Send/recv buff alias problem
This was less of a bug, and more of an unsafe code which was explicitly disallowed by MPICH (but not OpenMPI). Should be fixed now.
Per MPI standard
@LudwigBoess if you have time, could you test that this patch works?
Sorry for the late reply, the queue time on the machine I have access to are a nightmare. I was able to test the MPI on CPU and can't reproduce the issue with this version. The GPU MPI run has been in the queue for 5 days, so I can't test it.
Anyway I think we can consider the bug to be fixed!
@LudwigBoess no worries, let me know if it still persists, there is no rush about merging this.