Test backward compatibility with NumPy2.0

automl / DEHB

https://automl.github.io/DEHB/

Apache License 2.0

71 stars 16 forks source link

Test backward compatibility with NumPy2.0 #91

Closed Neeratyoy closed 3 months ago

Neeratyoy commented 3 months ago

For now we should limit Numpy dependency to be <2.0. Given we manipulate arrays and they form the crucial data structures for the algorithm, we should explicitly test what breaks and how significant changes are required for the new Numpy version.

eddiebergman commented 3 months ago

I highly doubt dehb would break, it more seems for libraries building on top of numpy, doing C ABI stuffs.

Of course worth checking, but it would take 5 minutes to test

Bronzila commented 3 months ago

Yup, I checked it yesterday and it seems like ConfigSpace breaks with Numpy 2.0 - That's also the reason why our tests crash at the moment.

eddiebergman commented 3 months ago

Yup, well aware. You can try this version of ConfigSpace which hopefully gets released soon: https://github.com/automl/ConfigSpace/pull/346

Bronzila commented 3 months ago

Just checked, whether DEHB itself is affected by numpy 2.0 changes, using ConfigSpaces PR version and DEHB does not crash. I still added the limitation to numpy <2.0 to pyproject.toml while we are waiting for the new ConfigSpace version. I will also create an issue for this in order to not forget about it.