It seems my conda configuration was the culprit in this weird installation of pytensors. My conda channel settings were
channels:
- conda-forge
- bioconda
- r
- defaults
auto_activate_base: false
channel_priority: flexible
Changing flexible to strict now asks to install pytensor 2.17 from conda-forge (rather than pytensor 2.13 from default and base 2.17 from forge). I will try reinstalling the environment now and running the code.
Done, installed it with strict and now environments are more uniform by default but there is still a speed difference between when doing pm.sample(random_seed=10, init=‘advi+adapt_diag’).