- The array object no longer defines `__array__`, and conversion to NumPy relies on the

buffer protocol, `__buffer__` instead. This change should be completely transparent to

users. Please report any issues this change causes however.

- Support for Python versions 3.9-3.11 was dropped. The minimum supported Python

versions is now 3.12.

- `asarray` no longer accepts nested sequences of arrays. This is consistent with the

standard, which only allows _a (possibly nested) sequence of Python scalars_. In most

cases, this requires changing `asarray(list_of_arrays)` to `stack(list_of_arrays)`.

Note that this effectively disallows code such as

`asarray([x[i] for i in range(x.shape[0]])` since indexing 1D arrays produces 0D

arrays, not python scalars.

- Fix fancy indexing in a mult-device setting. The indexed arrays and all indexer arrays

must be the same device. Otherwise, an error is raised.

- Make `finfo` and `iinfo` accept arrays or dtypes, as required by the standard.

- Make `roll` only accept integers and tuples for the `shift` argument.

- Make `reshape` only accept tuples for the `shape` argument.

- Make testing of array iteration compatible with Python 3.14.

The following users contributed to this release:

Lumir Balhar,

Evgeni Burovski,

Joren Hammudoglu

Tim Head,

Guido Imperiale,

Lucy Liu