Comment by dragontamer

With a 4B parameter model that probably ran through 8GBs of RAM multiple times to run.

At a certain point we should seriously talk about CUDA accelerating Postgres instead.