Stan math: Was ChainableStack::memalloc_ deprecated?

math
#1

A quick question, was stan::math::ChainableStack::memalloc_.free_all() deprecated? I’m trying to build an Rcpp package I’m developing as part of my PhD. It works fine on my one Linux machine, using the StanHeaders version 2.17.2 package. On my other Linux machine, using the latest StanHeaders package, version 2.18.0, it throws an error about memalloc_ not being a member of ChainableStack.

If it was deprecated, what is the current alternative to stan::math::ChainableStack::memalloc_.free_all()`? Or is this no longer required? I remember previously that if I omitted:
stan::math::recover_memory(); stan::math::ChainableStack::memalloc_.free_all();
after calculating gradients, then bad things would happen…

0 Likes

Errors getting my old R package that depends on the stan math library to build
#2

I think you can use like, ChainableStack::instance().memallloc should compile

1 Like

#3

Sorry for the late reply. The free_all() should be actually doing the memory freeing. We usually don’t do that until cleanup time given that our use case is MCMC or optimization and we wind up comput the same shaped log density function repeatedly, which can then reuse the allocated memory.

The two things to use ar recover_memory() if you only ever do top-level calls, and recover_memory_nested() if you’re doing nested autodiff (derivatives within derivatives, as we do, for example, in our ODE solvers).

The free_all() function should still work. Can you send us the compiler dump if it’s failing to compile?

0 Likes

#4

Apologies for my late reply! I’m a little swamped right now, and I’ve moved on to some more critical aspects of dissertation. For now, I’m just specifying that the user needs exactly version 2.17.2. But when I have some free time, I’ll try and see if I can’t slap together a minimal working example with a compiler dump. It won’t be this month though :-(

0 Likes