Repository navigation
Support for machine learning dtypes #372
Description
I am building a reference runtime for ONNX using this library, which needs to support data types commonly used in machine learning, e.g. bfloat16, float8*, float4, and lower bit ints like (u)int4 and (u)int2. These types exist in frameworks like jax, pytorch and numpy (via the ml_dtypes package). Is there a plan to support these datatypes?
Activity
ev-br commentedon Jan 1, 2026
No such plan yet, no.
What kind of support you're after? This
-compatlibrary scope is to smooth over small incompatibilities between array libraries and the Array API spec. The ML dtypes are not in the spec, thisarray-api-compatdoes not know about them; this is similar to other array lib features that are beyond the spec: the-compatversion does not block them, and leaves them to an array library itself.In the ONNX based land specifically,
ndonnxhas the Array API support in the main namespace, so maybe start with howndonnxhandles these ML dtypes.