I’ve edited the post to add some notes, let me know if it needs any more!
@rok_cesnovar this is good to go for the blog, thanks everyone!
3 posts were split to a new topic: User defined gradients updates
I’m not going to be able to attend next week’s meeting, but I only have two real updates:
- Complex matrices/vectors MVP (Complex containers by WardBrian · Pull Request #1133 · stan-dev/stanc3 · GitHub) is passing tests and should be merged in the next couple weeks if nothing major comes up
- I posted a design doc about language annotations (Java-style @annotations for user-guided optimization by WardBrian · Pull Request #45 · stan-dev/design-docs · GitHub). Hopefully @rok_cesnovar can join to discuss these
The monthly language meeting is tomorrow (Thursday) at 10 am NY time. If anyone would like a Google Calendar invite or Zoom link, please let me know either here or via email.
Looking like I’m not going to be able to make it either, please let me know if something comes up that I can help with
Monthly language meeting is again tomorrow (Thu—April 14) at 10 am NYC time. Please let me know if you need an invite to the Zoom meeting.
Notes from today’s meeting
@louis-mandel had some thoughts on the standard library refactor (see stanc3#1182, stanc3#1184).
He brought up a few advantages to the first approach which I had not considered, namely that there are very valid uses for wanting multiple libraries in the same compiler instance - his numpyro/pyro project lets you choose between those two backends with a flag, for example. @stevebronder and I are going to consider the original PR which would allow this more.
I brought up two design docs which have received positive feedback and seem ready for formal review, Tuples (which @stevebronder said he would look at) and Annotations (which @Bob_Carpenter said he would)
With the addition of complex vector types, we want to define signatures like
abs(complex_vector) => vector. To avoid unnecessary promotions, we should also vectorize the other signatures of
abs. This leaves
fabsas redundant, and we will be deprecating it. We will probably do something similar to
fmaxin the future as well.