Any thoughts on possibility of running the generation step on apple silicon via MPS instead of renting a cuda pod?

Sorry, I did not try it. Its worth a shot, I guess.