Mojo as a modern backend for PyMC?

[CReizner]

I know there's nothing better than a complete rewrite (joke!), but I'd like to throw out the idea of considering Modular's new pythonic language Mojo as the next foundation for PyMC. Mojo is made for AI applications and already supports tensor operations and computation on a wide variety of hardware, including nvidia graphics cards (without CUDA). Since bayese statistics will probably always be computationally intensive, perhaps it would be a good option for the future? At the moment, Mojo can integrate Python libraries very easily and their plan is to make it work the other way around. Maybe this is the possibility for you to make PyMC even better and to become more independent from C(++) libraries? What do you think?

[ricardoV94]

If anything it could be a backend for PyTensor, not PyMC. We are deeply intertwined with the computational graph rewriting and manipulations we do via PyTensor before we actually compile to some performant backend (we use pytensor not just for performance, but for the ability to transform computational graphs for different steps of the bayesian workflow).

So the question for someone interested is wehther mojo would fit as one of the backends of PyTensor, and/or be used for the graph representation/manipulation aspects.

Otherwise, without touching PyTensor, if mojo can replace the python part of the samplers, and still call / integrate well with PyTensor that could be nice.

I doubt we will take that as a goal for our project, but would be always curious about people playing with new technologies and reporting back whether they seem promising / a good fit.