switch from spectrum mpi to openmpi

[den-run-ai]

Do you provide any documentation about switching from spectrum mpi to openmpi installed in powerai 1.6:

https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/

[bnemanich]

Hi,

It is not possible to use the DDL, PyTorch or SnapML packages with OpenMPI, they require Spectrum MPI.

If you would like to use other PowerAI packages with OpenMPI, you will have to install OpenMPI on your system and create a conda environment that does not include either DDL, PyTorch or SnapML.

[den-run-ai]

Why do you provide openmpi in that powerai 1.6 list from the link above? If I try to install it via conda, then it is not available.

[bnemanich]

For the x86 version of PowerAI, OpenMPI is used in place of Spectrum MPI, since there is not Spectrum MPI available on x86.

So, on the channel above you are seeing OpenMPI listed for x86. If you click on linux-ppc64le on the above link, you will only see spectrum-mpi.

[den-run-ai]

@bnemanich do you provide instructions how to build Pytorch compiled with SMPI 10.3 or CUDA-Aware OpenMPI 4 instead of IBM DDL? Right now Pytorch in PowerAI channel is compiled with IBM DDL.

[den-run-ai]

cc @hartb for example here is an issue:

pytorch/pytorch#31438

[qhaas]

For the x86 version of PowerAI, OpenMPI is used in place of Spectrum MPI, since there is not Spectrum MPI available on x86.

It would be nice if IBM provided a ppc64le version of pytorch in their conda channel or as a whl that didn't have dependencies on Spectrum MPI and used the unencumbered OpenMPI like on x86-64. It is quite tricky to get one's Spectrum MPI license playing nicely inside containers, see #228