Extreme Training efficiency?

#4
by aaronday3 - opened

Hi there, great model!

The thing that caught my attention is that you trained a 64 rank qlora for a 70B model on 4x4090s which would have around 96GB of VRAM.

This is extremely efficient and I was wondering if it's possible to do this in axolotl or whether qlora-pipe has special vram optimizations that axolotl doesn't. Is it possible to do this in axolotl?

I reviewed the config you posted in the other discussion and I have a question, did you make qlora-pipe specifically for VRAM efficiency? Is it more efficient than axolotl?

Sign up or log in to comment