Sebastian Berg

> In the long run we'll ensure CuPy supports all kinds of possible hardware-accelerated narrow precision types Hmmm, yeah. Unlike structured dtypes, we don't need to worry about `scalars`/`dtypes` confusion...

I can still reproduce this with current Cython releases. More importantly, here is a nicely packaged reproducer (using `numpy>=2`: ``` import numpy numpy.random.RandomState.standard_normal("not a randomstate", 3) ``` It may crash,...

Hmmm, after profiling on the C/Cython level, I am not sure that there is a single big overhead here (this is not a custom elementwise kernel after all). Things may...