GitLab

Softcap isn't in the whitepaper/implementation for the language model so we should remove it. There is no discernible difference in output with it removed.

Currently we are using positions, which are relative to a
sequence and may not be unique.