Fine-tune LLMs using torchtune

[pparshakov]

Got same error. RTX 3090, torch.cuda.is_bf16_supported() gives True.