nice writeup, but i feel that for most people, the software side of training models should be more interesting and accessible. for one, "full" gpu utilization,

nice writeup, but i feel that for most people, the software side of training models should be more interesting and accessible. for one, "full" gpu utilization, one or many, remains an open topic in training workflows. spending efforts towards that, while renting from cloud, is a more accessible and fruitful to me than to finetune for marginal improvements. this course was a nice source of inspiration - https://efficientml.ai/ https://efficientml.ai/ - and i highly recommend looking into this to see what to do next with whatever hardware you have to work with.