Quantization?

#7
by andrewdalpino - opened

Hello ESM community,

First off, thank you for your contributions to the field of Genetics and Biology more broadly. I'm currently using ESM3 as well as a fine-tuned ESMC model as components of a "drug design" agent. I was able to do quantization-aware fine-tuning for ESMC but I'm wondering if anyone can speak to the ESM3 family of models in terms of ...

  1. Have there been any validations done on various post-training quantization strategies? ex. int4w vs int8w.
  2. Was ESM3 training using QAT?

Any additional comments regarding ESM3 in the context of quantization would be helpful as well.

Thank you!

Andrew

Sign up or log in to comment