Planning the next CmdStanPy release #690
Description
I think we should try to have a new cmdstanpy release in the next couple weeks. Depending on some choices we make this will either be 1.2.0 or 2.0.0 (discussion below).
Changes so far:
New Features:
- Allow git:TAG as a version in install_cmdstan. Allow git:TAG as a version in install_cmdstan #679
- [Stan 2.32] Laplace method and other changes. [Stan 2.32] Laplace method and other changes #669
- Add "chain__", "iter__", and "draw__" columns to draws_pd. Add "chain__", "iter__", and "draw__" columns to draws_pd #677
- Allow multiple dicts in inits, fix multichain. Allow multiple dicts in inits, fix multichain #688
- [Stan 2.33] Move IO munging to external package, refactors. Stan 2.33: Move IO munging to external package, refactors #681
- [Stan 2.33] Add Pathfinder. [Stan 2.33] Pathfinder #686
Fixes and Minor Changes:
- Correctly fetch ppc64el CmdStan. Correctly fetch ppc64el CmdStan #661
- Add version bounds on standalone GQ for non-HMC . Add version bounds on standalone GQ for non-HMC #657
- Fix: escape spaces in call to make. Fix: escape spaces in call to make #662
- Accept np.floating and np.integer for arguments. Accept np.floating and np.integer for arguments #663
- Only check version URL after determining if it was already installed. Only check version URL after determining if it was already installed #673
- Update documentation on external C++. Update documentation on external C++ #675
- Handle tilde
~in model paths.. Handle tilde~in model paths. #671 - Doc: spelling fixes. Doc: spelling fixes #678
- Fix handling of include-paths in compiler options. Fix handling of
include-pathsin compiler options. #689
The only reasons this would need to be 2.0 are in #681. The return type of CmdStanMLE.stan_variable changed from sometimes-float-sometimes-ndarray to always ndarray, and the meaning of CmdStanVB.stan_variable changed from returning the mean approximation to the draws from the variational approximation (see #636).
We could patch around/partially revert these two changes and instead emit warnings, release as 1.2.0, and then plan a 2.0 which would make the changes and remove anything else marked as deprecated. Or, we could just release what is currently merged as 2.0.
In either case, it is also worth doing a once-over on the API and seeing if anything else should be deprecated and removed. There was some discussion in #689 of the after-construction compilation being an odd feature of the model class, and discussion in #684 of various arguments with different names that all mean draws.