Numpy Serialization is not Unique for FP8 dtypes

[jon-chuang]

import jax.numpy as jnp
import numpy as np

>>> np.arange(10).astype(ml_dtypes.float8_e4m3b11fnuz).dtype.str
'<V1'
>>> np.arange(10).astype(ml_dtypes.float8_e4m3fn).dtype.str
'<V1'
>>> np.arange(10).astype(ml_dtypes.float8_e4m3).dtype.str
'<V1'
>>> np.asarray(jnp.arange(10).astype(jnp.float8_e3m4)).dtype.str
'<V1'

I just want to point out how big of a footgun this is, and it seems fundamentally misguided by mapping all the above dtypes to have the same kNpyDescrKind:

ml_dtypes/ml_dtypes/_src/dtypes.cc

Line 87 in d44671c

static constexpr char kNpyDescrKind = 'V'; // Void

Though I doubt we would want to change this now because of backwards compatibility.

[jakevdp]

Thanks for the report! This is a known issue, but unfortunately not something that can be addressed without fundamental changes to NumPy itself.