Bump PyTensor dependency

[ricardoV94]

Bumps PyTensor dependency to 2.14.1

This PyTensor release reverts dynamic broadcasting in Elemwise. Code that started to rely on it will fail, hence why it is marked as a major PyMC release as well.

We also keep shape dependencies in the original graph, so some logic that expected eager Constants had to be tweaked.

Finally, no more automatic support shape inference for Multivariate RVs. The thing we had going on wasn't robust and made errors cryptic in the cases where it didn't apply. This includes dropping support for Multivariate CustomDist not implemented with via dist

---

📚 Documentation preview 📚: https://pymc--6830.org.readthedocs.build/en/6830/

[codecov]

Codecov Report

Merging #6830 (845c8bb) into main (e3961fc) will decrease coverage by 0.02%.
The diff coverage is 95.00%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6830      +/-   ##
==========================================
- Coverage   92.05%   92.03%   -0.02%     
==========================================
  Files          96       96              
  Lines       16373    16378       +5     
==========================================
+ Hits        15072    15074       +2     
- Misses       1301     1304       +3     

Files Changed	Coverage Δ
pymc/pytensorf.py	92.75% <ø> (ø)
pymc/variational/approximations.py	90.54% <50.00%> (+0.04%)	⬆️
pymc/variational/opvi.py	87.57% <88.23%> (+0.09%)	⬆️
pymc/distributions/distribution.py	96.35% <100.00%> (-0.72%)	⬇️
pymc/distributions/multivariate.py	92.22% <100.00%> (ø)
pymc/gp/cov.py	98.00% <100.00%> (+<0.01%)	⬆️
pymc/logprob/mixture.py	96.92% <100.00%> (+0.02%)	⬆️
pymc/logprob/order.py	97.87% <100.00%> (+0.04%)	⬆️
pymc/math.py	69.15% <100.00%> (-1.47%)	⬇️
pymc/model_graph.py	80.43% <100.00%> (+0.32%)	⬆️
... and 1 more

... and 1 file with indirect coverage changes

[ricardoV94]

Had to pin arviz from above, as it was installing on the python 3.8 and failing

[twiecki]

Let's just drop 3.9 too.

[ricardoV94]

Let's just drop 3.9 too.

3.9 is supported by Numpy until next January. I think Oriol meant 3.8 in the comment he left, if that's what led you to say that?

[OriolAbril]

I did mean 3.8, sorry

[ricardoV94]

Last PyTensor release broke something on VI @ferrine.

https://github.com/pymc-devs/pymc/actions/runs/5585372543/jobs/10208039570?pr=6830

My guess it has to do with this one:

• Allow rebuilding graphs when output type depends on input values pytensor#280

But could also be an interaction with this one:

• Use static-only broadcasting rules to compute shape of broadcasting pytensor#345

Looking at the step_fn I see one extra NormalRV out of nowhere. And the first evaluation doesn't seem deterministic anymore either, maybe because the replacements are introducing a new RV that didn't use to be introduced (and which is not properly seeded by the logic in compile_pymc)?

Any chance you can take a look?

[ricardoV94]

Took 3 releases, but CI is passing! :)

[ferrine]

I'm good on the vi side, anything else?

[ricardoV94]

Thanks @ferrine! Also with the help debugging!

I requested another review for the non-VI changes. Will wait a bit for feedback before merging.

[michaelosthege]

major as in PyMC v6 ?

I'm just on the phone until Sunday, so I can't really check the diff. (Don't wait for mel

However, I'd like to request an example of a line of code that will fail and how to fix it to be included in the release notes.

[ricardoV94]

major as in PyMC v6 ?

No, major as in PyMC v5.X.0

We are not doing "major, major" releases unless the PyMC API / dependencies change drastically.
I don't have a better name for these, so I've been calling them major. Apologies for the confusion

However, I'd like to request an example of a line of code that will fail and how to fix it to be included in the release notes.

We are reintroducing static broadcasting-only again with the PyTensor bump, so for example broadcasting with MutableData("x", np.ones(1, 10)) will fail, unless you write MutableData("x", np.ones(1, 10), shape=(1, None)), or add a specify_shape after the fact.

All other needed changes were internally-only and shouldn't affect users. Mostly, I imagine it will break custom defined MultivariateRVs that relied on the old automatic shape inference..