facebook-opt-125m-qcqa-ub-6-best-for-KV-cache/README.md

license, language, library_name

license	language	library_name
mit	en	transformers

This is a QCQA version of the original model facebook/opt-125m. In this version, the original MHA architecture is preserved but instead of having a single K/V head, different K/V heads corresponding to the same group have the same mean-pooled K or V values. It has upto 6 groups of KV heads per layer instead of original 12 KV heads in the MHA implementation. This implementation is supposed to more efficient than corresponding GQA one.