Within-Chain parallelization & Seed

Modeling
1

Hi everyone,

I’m using the within chain parallelization (with the function reduce-sum and partial-sum), and I was wondering if we could specify a seed to obtain the same results from a run to another run of the same model, please ?

Indeed, I’m able to specify a seed in the r part and the stan part of my model, leading to the same results from a run to another run of the same model when I don’t use the within chain parallelization.

However, when I use the within chain parallelization with a seed in the r part and the stan part, I don’t obtain the same results from a run to another run of the same exact model (with the same exact data also). So I was wondering If there is a way to specify a seed for stan to perform the same within chain parallelization each time?

Thank you !

2

Using threading makes things not easily reproducible in an exact way. You may get exact reproducibility if you use the static version of reduce_sum. This version chunks the large sum always in the same packets…and hence the choice of the grainsize is more important to choose well for good performance.

3

Thank you for the reply!

How could I tell stan that I want to use the static version of reduce sum?

For now, I was using a grainsize of 1, which seemed to be the one with the best performance.

4

Please have a look at the Stan functions manual.

5

I’m guessing that @wds15 means the version without threading. To do that, you need to change the compiler flags to not use threading I believe. But I don’t know how to do that. @WardBrian or @stevebronder should know.

6

I believe he meant the reduce_sum_static function

7

Yes, that’s what I meant.

8

Thanks—I didn’t even know there was such a function. Sorry for confusing matters!