I am very much wishing for the same. I have used T0_3B (https://github.com/bigscience-workshop/t-zero) on my M40 with 24GB of vRAM. It was fun to get it working, but its not remotely comparable to GPT-3 (instructGPT).