Lately i been trying out the full bayesian approach and i have just come to a bottleneck, but lets state the good things first, by using the jax backend and sample_blackjax_nuts i got the sampling time down by *10 making it possible for me to go fully bayesian in the modelselection part. Reading through some of the awesome docs out there on PSIS-loo i decided that it is fine in my scenario but that i use k-fold cross cv in case of a lot of strange pareto-k values for assumingly good models(reloo scheme can come to the rescue). Now to the bottleneck, in my task, i have an downstream optimization problem which needs to evaluate the predictor ALOT, currently i am setting new data(proposed by the optimizer) by set_data and subsequent sample_Posterior_predictive but this is way too slow(3-4 sec per evaluation), do you have any ideas on how to tackle this(using v5), the obvious one being downsizing samplesize etc… thining out etc… but maybe there is something else that i am not aware of?