The issue in the CI does not originate from scipy, but instead from incompatibility of tensorflow probability to numpy 2.0. At first, I could not reproduce the issue locally, because locally, I installed using the pmc specs, and pymc currently pins to numpy<2.
The current solution is to pin numpy in liesel to <2.0, too, until the downstream issue is addressed. There's already an issue in the tensorflow probability repo:
See this comment:
https://github.com/liesel-devs/liesel/pull/198#issuecomment-2248741496
The issue in the CI does not originate from scipy, but instead from incompatibility of tensorflow probability to numpy 2.0. At first, I could not reproduce the issue locally, because locally, I installed using the
pmc
specs, and pymc currently pins to numpy<2.The current solution is to pin numpy in liesel to <2.0, too, until the downstream issue is addressed. There's already an issue in the tensorflow probability repo:
Once this is addressed, we can unpin numpy, since pymc correctly handles its dependency on numpy itself.
Originally posted by @jobrachem in https://github.com/liesel-devs/liesel/issues/198#issuecomment-2248741496