if torch.cuda.get_device_capability()[0] >= 8:    !pip install -qqq flash-attn    attn_implementation = "flash_attention_2"    torch_dtype = torch.bfloat16else:    attn_implementation = "eager"    torch_dtype = torch.float16