I’m going to say it’s maybe half way done. Completion involves parsing the Stan language into the AST, translating the AST into the Middle Intermediate Representation (MIR), and generating code for that. On top of that we’ll need to allocate some time for figuring out distribution for RStan/CRAN, and do a bit of work for PyStan. @Matthijs has the entire language parsing (with 5-8 bugs fixed, something like that) into an AST that is being finalized today. The translation to MIR is probably 95% done, and the code generation is something like 40% done.
I think it might be ready in ~3 months or so.
I think we’ll probably end up releasing another version with the existing C++ compiler for the GPU and threading stuff if nothing else.
I don’t, however, think it’s worth spending Columbia or NumFOCUS money fixing the bugs that we’ve only discovered through the extensive test suite we’re (well, mostly @Matthijs) developing for the new compiler, unless a user chimes in that they have experienced this.