I think you can and should use other structures than vector[memory_view]. For instance, using vector[vector] is possible and such structure is coercible to nadarray[nadarray] if needed.

But I would focus our energy later on this PR as scikit-learn#24076 might be structured again with round of reviews.

Just linking a comment for when you come back to this cython/cython#4966 (comment). I proposed using a list of Python objects wrapping memoryviews.

I've only quickly skimmed this PR so I don't know if it actually does what you need though, so feel free to ignore if it isn't relevant.

Thanks @da-woods for this proposal.

I read this proposal in the original issue. Here, we might want to have instructions run as much as possible (or even solely) in nogil context for optimal performances. In this case, I do not think it's possible to use Python lists, isn't it?

Here, we might want to have instructions run as much as possible (or even solely) in nogil context for optimal performances. In this case, I do not think it's possible to use Python lists, isn't it?

No I agree. This is the big limitation of that suggestion.

@Micky774 Is this implementation in a working state now?