Any plan to support float16 ?

[Hzc492]

Hi,
Thanks for the great work on this library!

I'm currently using it to encode float types of various widths, and I noticed that float16 is not currently supported.

My questions:

1. Are there any plans to support float16 in the future?

2. If not currently planned, could you give some guidance on: What would be the key components that need to be modified or extended to support float16?

Thanks a lot in advance!

[azimafroozeh]

Hi,
Thanks for your kind words!

We’re seeing increasing interest in reduced-precision formats such as bf16, float16, and even fp8 (E4M3 and E5M2) — related issue here — so we definitely want to extend ALP to support these types.

However, supporting these formats will require benchmarking to effectively tune ALP’s parameters.

Would you be able to share a dataset, or point us to a real-world dataset similar to yours, that we could use to benchmark float16 support in ALP?

[Hzc492]

Our main usage scenario is storing inference models. We convert the models from float32 to float16 for storage and hope to further encode them using ALP. Therefore, most inference models can serve as benchmarks for testing. For example, the models from https://huggingface.co/. You can manually convert them to float16.

Thanks a lot！

[Hzc492]

Hi @azimafroozeh, maybe this dataset can be considered :)
https://huggingface.co/datasets/saget-antoine/francecrops-float16

[azimafroozeh]

Thanks for sharing the dataset! I'll look into it and get back to you.