Closed jakevdp closed 8 months ago
The checks for already registered types were added in the precursor to ml_dtypes, when both JAX and tensorflow registered their own copies of bfloat16. Since that is no longer a concern, we can remove this logic.
There's still another followup here, to allocate dtypes statically rather than on the heap. I plan to do that in a later PR.
The checks for already registered types were added in the precursor to ml_dtypes, when both JAX and tensorflow registered their own copies of bfloat16. Since that is no longer a concern, we can remove this logic.