You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Could you provide scripts for training larger models such as Llama-3.1-70B-Instruct.
It seems that only data parallel like ZeRO is enough for 8 GPUs, since only AttnGates are trained. But I do not know whether my calculation is right. Did you use model parallel during training?