Should there be a integral and floating ABCs?

[NeilGirdhar]

That is, should np.integral, np.floating, and np.complexfloating base classes be available in the Array API? This would allow other packages to create their own subtypes. Jax for example has additional floating dtypes.

[rgommers]

Hi @NeilGirdhar, thanks for asking. Whether dtypes should have a class hierarchy and how to check whether a dtype is of a certain kind was pretty extensively discussed in:

• RFC: add data type inspection utilities to the array API specification #425
• Dtype / checking standardization #152
• also relevant: ENH: add a canonical way to determine if dtype is integer, floating point or complex numpy/numpy#17325

tl;dr not many libraries have a class hierarchy like numpy does, and it doesn't seem either necessary or a good idea to impose one in the standard

This would allow other packages to create their own subtypes. Jax for example has additional floating dtypes.

There should be no issue with JAX including additional dtypes in what isdtype does. Or have them work with dtype= keywords. Do you see a concrete problem with the current design?

[NeilGirdhar]

Thanks for the links. I've read over the discussion now.

There should be no issue with JAX including additional dtypes in what isdtype does. Or have them work with dtype= keywords. Do you see a concrete problem with the current design?

How are you supposed to do type annotations with the current design? How are type checkers supposed to verify that an array has a floating dtype for example? For example, https://github.com/NeilGirdhar/tjax/blob/main/tjax/_src/annotations.py#L24

[NeilGirdhar]

I'm going to comment in that thread.