Filippo Vicentini

Ah, I see! By the way, your way of encoding `Nbosons=9` as the local cutoff is... well, not wrong, but... it's completely different from his point of view that there...

I think that is the difference between `pickle` and `cloudpickle`. Using the former should avoid the problem. EDIT: yeah, but this won't work....

I strongly suggest merging this as well. Getting the version wrong becomes impossible after this. BTW, This will break the current conda-forge recipe, but can be easily fixed.

It's here https://github.com/conda-forge/flax-feedstock but don't worry, I'll do it myself.

I think what he means is that after you tag a commit with a certain version (eg 0.7.1) you should not delete this tag and tag a new commit with...

(I am the author of `mpi4jax` so I know that one pretty well 😄 ) Thanks for the explanation, I see the issue now. The reason I'm asking that is...

> Yes, in essence you are asking to replace the collectives emitted by XLA internally with calls to mpi or something similar. This is a bit different to mpi4jax where...

@hawkinsp did anything change on your end about this recently, or you're still not really planning on supporting this-?

@jon-chuang thank you for looking into this. It's something that would greatly benefit many people including me... As per your question, If I understand your question correctly, you want to...

Also, relatively relevant, how are you going to implement this? Ideally this could be a plugin to XLA (do those even exist?) that depends on MPI. Not so far off...