wds15

OK... I was just worried about the rare need for binary compatibility in the model class, but it sounds as if we are good.

more memory is needed, for sure... don't know by heart.

He made the suggestion in the parallel HMC thread... and I do agree. I do want to look into that - imagine you fire up 6 chains and pool the...

Why can’t we have a multi chain warmup which are terminated once done doing warmup. Then the CmdStan or whatever else interface fires up multiple chains with the so obtained...

@yizhang-yiz Our dynamic HMC implementation goes randomly forward and backward in time. Right now that is done serially. The speculative HMC allows to go simultaneous forward and backward in time....

And the coolest thing about the threaded warmup is that no changes to Stan-math reverse mode AD is needed. Everything will work with the existing infrastructure... though figuring out how...

I agree. The TBB has the concept of task_arenas which is a way to fence work from one another. My intuition is that we should manage these task_arenas from within...

I will start simple - which means by chain. If you do by draw then you run instantly into exact reprocubility hiccups unless you sort the data as you accumulate...

... and we need to be careful to keep things deterministic... I don't want that the workload on the machine determines the exact numerical output. It's a non-trivial restriction in...

Bump. @betanalpha this looks all good to me. The intent to error out when the chosen initial leads to a nan for the log-likelihood **or** it's gradient is sensible, since...