papersnake·4 年前·議論Have you tested this on big models involving multi-gpu communication, or any plans?