In RStan, information about leapfrog time and, more importantly, initialization issues such as dimensional mismatch in data input are well reported.
This is not the case in cmdstanR where the program simply crashes silently. This prolongs debugging time as we basically have little clue what is going on.
// I should I have fill an issue on github but I’m not sure which repo should I do, therefore I post here.
Can you post a lightweight reproducible example?
I’m not an R user, but I have used PyStan and then mostly CmdStan and cmdstanpy, but never noticed any significant difference between them with respect to that.
If you are noticing it, a good way would be to run the model you’re having trouble with in each interface at show how they are different. It can also be the case that you are getting a different, system error, and not seeing the Stan errors correctly.
Okay. I found out that this only happens in very big model where cmdstanr just crashes (or OOM-killed) and rstan doesn’t. That’s why the error only exist in the latter. I will keep investigating and close this for now.