[Question] What combination of filters is equivalent to a bfloat16?

[alxmrs]

My understanding is that Zarr is capable of storing ML model weights in distributed storage. Reading the documentation on filters, it seems like numcodecs already has all the requisite interfaces to support reading and writing bfloat16 datatypes (or really, any other crazy data type researchers find useful for modeling). I'm curious if anyone reading this knows the proper recipe of [codec/filter/etc]s to represent bfloat16 dtypes.

In think in practice, folks might use this library along with AsType() during encoding/decoding: https://github.com/jax-ml/ml_dtypes

[TomNicholas]

See also zarr-developers/zarr-python#2656

[d-v-b]

I think we should just support bfloat16 as its own dtype, along with any other dtype that's useful to people.

[alxmrs]

Thanks, that makes sense to me.