Pathfinder: unexpected behavior when num_draws < num_elbo_draws

[WardBrian]

Summary:

Reported by @avehtari:

Requesting the number of draws to be some small number and not also changing num_elbo_draws leads to a CSV file of unexpected length.
This is only the case when running single-path pathfinder or setting psis_resample=False. The number of samples returned seems to be max(num_paths*num_draws, num_paths*num_elbo_draws)

Reproducible Steps:

How many draws would you expect the following to return?

./bernoulli data file=bernoulli.data.json pathfinder \
num_psis_draws=100 num_draws=2 psis_resample=0 num_elbo_draws=3 num_paths=2

Our docs suggest the answer should be 2×2 = 4, but it is actually 2×3 = 6.

As an aside, it might be useful if cmdstan was issuing a warning when psis_resample is 0 but a non-default value was given for num_psis_draws.

Current Version:

v2.34.1