Interesting, I wasn't aware of #15097. Since it doesn't look like that pull request is ready to be merged, I think that we should merge this smaller fix first.

It'd b cleaner to explicitly declare the dtype. That's what we do in sklearn/ensemble/_hist_gradient_boosting/common.pyx for e.g. HISTOGRAM_DTYPE and hist_struct.

I'm fine with the proposed way, it's probably not worth nitpicking.

Can you please make sure that the resulting dtype is properly aligned to reflect the struct? The struct will be aligned, but numpy dtypes aren't by default.

I initially sent https://github.com/scikit-learn/scikit-learn/pull/16140/files to clean up the way that field offsets are calculated, but then decided that it would be cleaner (and harder to screw up) to use np.asarray as the comment suggested.

And yes, the np.asarray method is guaranteed to get the correct alignment.

Thanks @alexhenrie !