pymc
pymc copied to clipboard
pymc-devs
Rather different compilation times when PyMC is installed from pip and from conda-forge
Hello,
I am not sure if this is a bug, but perhaps a peculiar behavior that I'd like to report.
Over the past few weeks I noticed this: compiling a model takes anywhere between 20 and 80 seconds when PyMC is installed via conda-forge (most often in the 30-34 seconds range), and 1-2s when PyMC is installed via pip. In all cases I am using the experimental jax / numpyro sampler tho. I have not (yet checked) if this is also true when using the regular NUTS sampler.
The sampling time is practically identical - so no issue there.
I was just wondering if you are aware of this, and maybe it could point to a potential problem somewhere. I have done various checks, and there is no difference in the results between PyMC installed via pip and via conda-forge, apart from this rather large model compilation times.
In case it matters, the models I am compiling contains 2-4 Gaussian distributions, sometimes, a couple of Beta/Bernoulli distributions, and some deterministic ones. Nothing all that special or fancy in any way.
Cheers!
Versions and main components
- PyMC/PyMC3 Version: 4.1.4, 4.1.5
- Aesara/Theano Version: 2.8.2
- Python Version: 3.9
- Operating system: Linux
- How did you install PyMC/PyMC3: conda & pip
My guess is that any differences would come from the JAX compilation side. Would be useful to have an exact example so that someone else can try to replicate it locally.
Ok, let me see if I can provide an example over the next couple of days. Thanks!
Closing this for lack of info. Feel free to reopen / comment if there are new details so we can reproduce.