Sebastian Berg

Sorry. But in the end I suspect you guessed: I have a very firm opinion that any change that has enough fallout must go through deprecation. It matters whether the...

My opinion is just that I would be happy to merge at last a temporary revision or a deprecation if it makes your life or your users life easier. If...

> it seems that the issue might be within the numpy library. Right either NumPy or the underlying blas is presumably making an unaligned memory access. It would be helpful...

@glaubitz it is unclear to me why you don't see anything in the core dump. Maybe because there are many threads from OpenBLAS and it fails in one of them...

FWIW, I think this is correct. `NDIM` is zeroed currently, but since there is no shape or strides there is no way to use ndim at all, i.e. if `ndim...

Initially you said > passing `PyBUF_ND` is an suitable workaround And that is clearly always required. And we should consider this a Python documentation issue if you read the docs...

To summarize: I will merge a PR changing `->ndim` to 1 in this case and if pinged, chime in on discussions to change the Python docs. I do not agree...

This was intentional: https://numpy.org/devdocs/release/2.0.0-notes.html#np-unique-return-inverse-shape-for-multi-dimensional-inputs It is now compatible with `np.take_along_axes`. But we may have underestimated the impact of this choice, and whether it is even the right one, since previously...

Ah, one thing I missed and this does indeed improve, is that for `axis=None` returning this shape does simplify reconstructing the original array. Although, interestingly if paired with `np.take()` not...

@nabenabe0928 I forgot to mention that the work-around is to add a `.reshape(-1)` to the result, I think. And whatever happens, there is a small chance it could become a...