Multi-GPU training tries quantising the base model on every process at once

[bghira]

Reported from Reddit: https://www.reddit.com/r/StableDiffusion/comments/1epd0bc/comment/lhlnx5r/?context=3

Starting training with Quanto enabled results in each GPU process trying to quantise simultaneously, which hurts a lot when it comes to system memory use. Can run out of memory on a 250G VM with 8 GPUs.