Roadmap of new release?

[geniuspatrick]

It looks like you’ve made a lot of great improvements recently; are there any timelines for future releases that you can share?

[rwightman]

@geniuspatrick don't really have a specific time in mind yet, mostly gated by when I get a chance to run through some different training scenarios to fully verify additions and finish some regression checks of existing models. Most of the new stuff are model archs that don't have weights yet, and training code.

Major features that still need to get in before going through that process are the Mammut / Coca2 additions, and probably some generation (kv-cache) optimizations and tokenization cleanup.

I feel pypi release is not the highest priority right now given that most people doing training are probably cloning and/or doing editable installs anyways... once I do a 4.x pypi release everyone that has this as a dependency in various projects will start getting the new code and any regressions could cause some major issues...

If there are any major bugs that come up on 3.x I can still do a patch and release off that branch.