chamuyaye

> you can replace flash_atten with the normal attention in pytorch. Things will still work, despite that the training speed will be very slow. Can you give detailed steps, thank...