We can do a simplified computation of number of ( checkpoints + 1 ) * model size in fp32 and provide warning if there is a lack of disk space.
This will help manage expectations, as checkpoints do not provide proper error reporting when it runs out of disk space