Would this just be pm.sample_prior_predictive()?
My intuition is that time isn’t a great indicator of model correctness. A single level glm will run much faster than a hierarchical model even if it’s wrong, it’s just stealing degrees of freedom. I have simulated models with simulated data for which I know the exact dgp that still take 10+s/iteration. It may make more sense when comparing similar models with different covariates sets.
But I do like the idea of just sampling from the model and comparing the data it produces to your data. Faster way to rule things out if you have a slow model before doing a proper estimation.