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
To get a better precision, sometimes I have to force some layer like LayerNorm to be fp32 and other layers be fp16, in pytorch training, I could use torch.cuda.amp.autocast and convert layer input to fp32 to train, but in deepspeed how could I do this?
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
-
To get a better precision, sometimes I have to force some layer like LayerNorm to be fp32 and other layers be fp16, in pytorch training, I could use
torch.cuda.amp.autocast
and convert layer input to fp32 to train, but in deepspeed how could I do this?Beta Was this translation helpful? Give feedback.
All reactions