search

stan-dev / cmdstanr

CmdStanR: the R interface to CmdStan
https://mc-stan.org/cmdstanr/
Other
144 stars 63 forks source link

Unifying thread syntax across samplers #996

Open saudiwin opened 5 months ago

saudiwin commented 5 months ago

Is your feature request related to a problem? Please describe. This isn't a big deal, but now that we have multiple sampling algorithms, it would be nice if we could condense the syntax for specifying multi-threading. Currently, the sample method has threads_per_chain, pathfinder has num_threads, and laplace has threads. Especially as some of the utility of the variational methods is in using them as testing or initialization for MCMC, it would be nice if we could unify this syntax.

Describe the solution you'd like Perhaps just a single argument like num_threads or perhaps for compatibility with sample the others could all use threads_per_chain.

Describe alternatives you've considered Perhaps the other variational methods could also take threads_per_chain as an additional argument if threads or num_threads isn't specified.

Additional context This is mainly for using cmdstanr in a pre-compiled model context in which switching from one sampling method to another may be done for the same data/model.

karimn commented 5 months ago

This tripped me up recently. I switched from variational to pathfinder and was confused why my threads = 12 didnt' work anymore.

SteveBronder commented 5 months ago

I think we should have threads for all of them. The samplers should be using the multi-threading across chains available in cmdstan. So then the user just specifies how many chains they want and stan internally does thread management for within and across chain parallelization

SteveBronder commented 5 months ago

We have not done that yet because it is a pretty big rewrite of some internals of cmdstanr