Thanks for the quick review, Christian! 🙏

Thanks all! 🙏

For more context, perhaps this example makes the bug clearer?

from libc.stdint cimport uint8_t

cpdef Py_ssize_t test1(uint8_t[::] a) nogil:
    return a.strides[0]

cpdef Py_ssize_t test2(uint8_t[::1] a) nogil:
    return a.strides[0]

import numpy as np

a1 = np.arange(9, dtype="u1")
a2 = np.arange(27, dtype="u1")[::3]

test1(a1)  # 1
test1(a2)  # 3
test2(a1)  # 1
test2(a2)  # ValueError: ndarray is not C-contiguous

So leaving the 1 out meant the array did not have to be contiguous (even though the function assumed it was internally). The change here makes sure this is caught when passing the array in.

Good catch! I was unconscious of possibilities for no -continuous arrays!

There are plenty of memoryviews' definitions and declarations (including but not limited to functions and methods' signatures) that aren't specific enough in scikit-learn's codebase then.

I think we can open a meta-issue for referencing the problems and list such definitions and declarations.

I think we can open a meta-issue for referencing the problems and list such definitions and declarations.

Note that scikit-learn's codebase will use non-contiguous memoryviews when the input is not guaranteed to be contiguous.