Comment by papersnake

Have you tested this on big models involving multi-gpu communication, or any plans?