Open butlerpd opened 2 weeks ago
See https://hackmd.io/B_zU26JfSI-703Xz85Ie_Q?view and https://github.com/numpy/numpy/issues/26191 for ecosystem compatibility of major packages.
For the time being, if we pin numpy and leave everything else floating, pip should be able to pick compatible versions of these other dependencies.
As for when we unpin, it also looks like most things are already updated, or not really numeric, with some questions remaining about:
Periodictable needs a PR merging, and a release. It's not currently compatible
need to check the uncertainties package for compatibility - most other other dependencies seem like they might be updated?
also check model marketplace.
numpy 2.0
was released on Jun 16 and has a number of breaking changes such asnumpy.NaN
is no longer recognized (causes fatal error) and must be replaced withnumpy.nan
.requirements.txt
does not currently pin to a version so the short answer is probably to pinnumpy
to a version <2.0. However we should probably start working on compatibility? Question: How many dependencies will also have this problem?