xformAI/opt-6.7b-ub-16-gqa

Source: 🤗mit6.7B activated-other⚠ No SHA-256 weightsupdated 1,005d ago
or submit a torrent

This is a GQA 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 16 groups of KV heads per layer instead of original 32 KV heads in the MHA implementation.