HF exported models are always float32

[versae]

I've noticed that despite training on bfloat16 models and having c=bfloat16 in the model config when training, the exported model to the HuggingFace hub is always float32. Is there a way to have the exported models match whatever parameter I set as c in the model config?

[dlwh]

I mean, I think the default being the "param" dtype is probably how I want it (since bf16 is of course more lossy), but I think it makes sense to add a specific dtype argument to the export