[danielhanchen]

For experimentation and smaller models, single gpu is the way to go! Tbh I normally find most people to spend the majority of their time on datasets, training loss convergence issues etc!

But if its helpful I was thinking about spinning up a platform for something like that!