Yes, ArviZ functions take InferenceData or something that can be converted to InferenceData as an input. That means that ArviZ functions work when they get PyMC3 MultiTraces, PyStan fits… but they work by converting to InferenceData and using the resulting InferenceData. I can’t recommend strongly enough converting to InferenceData (both using az.from_pymc3 or return_inferencedata=True are fine) and then call ArviZ functions using an InferenceData object. And given that PyMC3 delegates stats and plotting since some versions back to ArviZ, pymc3.summary, pymc3.traceplot, pymc3.loo… are all ArviZ functions.
We recently updated the docs on the pymc3 website to directly link to ArviZ docs, but the internal conversion is probably not well explained anywhere. Using InferenceData will also have other advantages aside from this performance effect. You’ll get an html representation of your object, labeled dimensions and coordinates, automatic broadcasting… The radon notebook is a good example of integrating ArviZ into the workflow. Hopefully the rest of the documentation will also be updated progressively to avoid patterns such as calling ArviZ functions with multitraces.
Regarding the comment about loo.function. Also note that one of ArviZ’s goals is to ease sharing and reproducing the results. Having the inferencedata corresponding to a particular model should be enough to repeat the result analysis and exploration: plots, ppc checks, model comparison… This does not only affect ArviZ-PyMC3 but also PyStan, Pyro, even Turing in Julia, and hopefully inferencedata stored as netCDF will soon be compatible with the posterior R package too.
Supoorting the function as loo does is interesting to minimize memory usage in these cases but it’s not really compatible with sharing the netCDF with a colleague who uses Julia so they can repeat and extend the analysis or publishing the netCDF together with your paper so R users can reproduce your results to make sure their implementation is equivalent before using it in their own models.