how much GPU memory does it cost to SFT train a v5.1-30B model? To make it fit into a 80GB memory, how to set learnable modules?