Possible additon of quantization or Nfloat4 for lower vram (Same quality as fp8/bf16)

I tested nfloat4 quite a bit on onetrainer, and the results are basically the the same as sd scripts, but almost 9gb vram less.

I was wondering if its possible for you to implement it on your script?

Thank you!

The issue does not name files, tests, or an entry point. Start by comparing nfloat4 support in OneTrainer with the existing quantization paths in sd-scripts, then define how equivalent quality and the claimed VRAM reduction would be measured before implementation.

Written by the indexing model from the issue text.