Allow some chains in multi chain samplers to fail during initialization about stan HOT 3 CLOSED

This alongside your comment linking to this in CmdStanR about this being a disadvantage of the single-process multi-chain feature made me look up what we do in cmdstanpy, which is we do not return samples if any chain fails (even in the legacy one-to-one chain-to-process set up)

Assuming we keep that behavior, it makes no difference to us what the underlying implementation does, besides perhaps wasting time if some chains still run.

One question I suppose is what is the process return code of the process when only some chains fail to init?

from stan.

First, this is usually a bug in the Stan code. It's really common to miss the constraints required for a parameter, like not setting a scale parameter with a lower bound of zero. That gives you a 50-50 chance to initialize correctly. In those situations, we want to signal that the Stan code needs to be fixed.

Second, I don't think it'll be good for users to run a fraction of the chains they request. Maybe if some chains initialize, keep trying the other ones? But if a user wants 4 chains and they only get 3, they're going to have to run again.

In general, I think we should warn on bugs rather than try to cover them up.

The one place where init can be hard is in something like a high-dimensional regression. In those situations, I've found reducing the initial scale of initialization from (-2, 2) to a narrower interval gets rid of the problem.

from stan.

I'm leaning on the side of 'no' now that I'm thinking about this more. Bob I agree usually if I'm getting inits that bad idt running chains over and over is going to fix it and I probably need to think more about the model. Going to close this

from stan.