Stan Seibert

Yes, this is correct, and this is why I described this as a "research project". :) If someone figures this out, that would be a big step.

We don't have any special contacts with the Apple developer teams, but if someone out there does, please let us know. :) I should note that PyTorch is using Metal...

I think our best bet is figuring out how Julia did this: https://github.com/JuliaGPU/Metal.jl (I'm still not sure how it works, since it seems to delegate most of the work to...

This is a good idea (forgot that dataclasses were added in Python 3.7), though for some use cases the Python types will not be granular enough for Numba. In particular,...

That said, I wonder what happens if you try to define a dataclass using Numba types...

Oh sorry, I mean the other direction: Does using Numba types affect the usage of the dataclasses from Python? There is no way Numba currently would understand dataclasses, but it...

The first three failure are due to long deprecated NumPy features that were finally removed in NumPy 2.3: https://github.com/numpy/numpy/pull/28254 I think we can delete those tests. The functionality has been...

And looking at the final failing test, this seems to be a subtle issue about Numba's `typeof()` implementation: https://github.com/numba/numba/blob/6cc96d049d9a486fb03eef1f2479a792596eb4e6/numba/tests/test_types.py#L564-L571 This test is confirming that passing a NumPy record dtype through...

OK, talking with Siu, it looks like the final test failure is related to this deprecation: https://github.com/numpy/numpy/issues/25306 And the test isn't really testing anything important. (It is also > 9...

Correct, this is definitely a performance bug, although a long standing one, apparently.