Proposal: Make compiler optional and fall-back to numba or JAX

[twiecki]

Description

Installing PyTensor with a full C++ compiler chain is very fragile and difficult to maintain. We have alternative backends that are much easier to maintain. While it seems we can't yet switch to them by default, how about:

• We make a working c-compiler optional in the installation
• If no working C-compiler is found, we fall back to use the JAX or numba backend.

That way, people that can't get their compilers installed, still get fast PyTensor.

[ricardoV94]

The C-compiler is optional. The issue is some libraries are not linking correctly (I think), but this is hard to see until we try to compile an actual function. But that means it's not easy to know when to swap the default either.

[ricardoV94]

Also at this point we know neither numba nor jax support is good enough, so the default would have to be python-only

[twiecki]

What is "good enough"? Maybe we can have JAX as default and Python as fall-back?

[ricardoV94]

What is "good enough"? Maybe we can have JAX as default and Python as fall-back?

As in you'll break everything for 99% of the users who go through PyMC.

Don't forget it's not only pm.sample, which would also fail, but also prior/ posterior predictive which fail because we have dynamic shapes everytime someone uses coords