On Mon, 2025-11-24 at 10:33 -0800, Amelia Thurdekoos wrote:
> > ...could we at least try to use the bit width in the name such as
> > float16
> > as much as possible?
> 
> 
>  This is all excellent feedback, thank you! I agree, more explicit is
> better. Planning as best as we can with an eye to future clarity,
> seems
> reasonable.
> 
> 
> > The important part here seems to be pushing the goal of getting rid
> > of
> > npy_math, what is the plan for the rest of it here?  Just getting
> > rid
> > of it completely because it is obsolete?
> 
> 
> 
> If we introduce this now users will adopt it in 2-3 years (old NumPy
> > versions don't have it).
> > Now there is no guarantee that e.g. `float16_t` will be generally
> > supported (it's optional anyway) but if the situation around half
> > improves just a little, I wonder if by the time this is useful it
> > won't
> > already be completely obsolete, rather than just mostly.
> 
> 
> Thanks for the feedback Sebastian! I agree, the overarching goal is
> removing npy_math. This proposal's scope is intentionally limited to
> the
> float16/halffloat part as one self-contained step in that direction,
> rather
> than trying to solve the whole npy_math feature at once. I am not
> going to
> be removing any part of the library as part of this effort.
> 
> If float16 improves enough that this API becomes obsolete, or even
> less
> attractive, this effort still yields a net positive. It will assist
> with
> reducing the need to link against npy_math.
> 


Yeah, nice and I agree with Nathan that it is tedious for e.g. DType
casts. ml_dtypes uses the Eigen for this purpose CuPy vendors [1].

If `npy_math` wouldn't be deleted it would be nice to have a thought on
what happens with it (since we could probably use the same thing for
half).  But I think the hope is to just delete and I don't want to
block this if this seems like the main blocker.

However, it still seems to me like only the conversions are (for now?)
truly useful. And I also still wonder if header-only may not be just as
well and simpler for all sides?
I would also be happy to consider to only keep the conversion functions
and ask users who need serious math to look elsewhere, because for now
I still suspect it's not hard to find things elsewhere (e.g. Eigen).
(Yes, I am assuming that would throw at least close to zero
persons/projects under the bus if this starts missing.)

- Sebastian

[1] For now CuPy never compiled against NumPy, if it does there are
better replacements, i.e. CuPy wouldn't need this functionality.


> - Amelia
> _______________________________________________
> NumPy-Discussion mailing list -- [email protected]
> To unsubscribe send an email to [email protected]
> https://mail.python.org/mailman3//lists/numpy-discussion.python.org
> Member address: [email protected]
_______________________________________________
NumPy-Discussion mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3//lists/numpy-discussion.python.org
Member address: [email protected]

Reply via email to