@Bob_Carpenter, there have been two bugs that were uncovered recently:
matrix_exp_multiply() can seg fault. There’s been a fix applied, which is slow, but won’t seg fault. #1073
- There’s a major bug in threading. Depending on the number of threads and number of shards, the model’s log prob function isn’t computed correctly. #1075
I think once these two bugs have been squashed, we should release v2.19.0. Thoughts?
Yup… makes sense. The
map_rect threading code needs a released fix ASAP as quite a few people seem to use this feature. We should have the fix for this done this weekend.
(it’s a pity to see that GPU things are not yet in 2.19 then, but I guess we can release a 2.20 once it’s ready)
There are also other nice things in develop right now (fixed makefiles for MPI, faster ODE integration are things I recall).
Yeah that’s my bad, work got a little hectic and Rok is finishing up some stuff for the latest PR. What is the timeline for 2.19? I don’t see the last PR for the GPU stuff (rev of Cholesky) taking longer than January 10th or so
the timeline will probably be discussed at the next stan meeting… which I can’t join. Right now @Krzysztof_Sakrejda is preparing a fix for the threading issue and if I were to decide it, we should release before xmas, but I don’t know what others think. So maybe @syclik can bring this up on Thursday and get consensus?
EDIT: And don’t worry about timing… the work you contribute is amazing and as I understood on top of your usual work.
I’m trying to get a model in shape for next week so at best I’ll get a few spare minutes to modify the testing later this week. …