[Question] What prevents DefaultCommunicator from working with multi-node MPI

[XapaJIaMnu]

Hey,

I have been experimenting with some CPU training (without CUDA/NCCL) and I run into this ABORT statement.
https://github.com/marian-nmt/marian-dev/blob/master/src/training/communicator.h#L128

However, I don't see immediately what implementation part is missing in order to get full MPI support. I tried commenting out the ABORT and it seems that it works. Are gradients not exchanged? Git blame shows @frankseide put it there?

Could you guys let me know what part I should add to properly fix this issue?

Cheers,

Nick

[emjotde]

We talked per e-mail, but basically you are welcome to own that part.