Models arbitrarily large running and training by streaming off of disk, allowing scaling using a fixed amount of VRAM and running on machines with little VRAM. This is a 40b parameter model trained at regular precision on a card with just 96gb of memory
