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
I have a problem where I need a scalar weight tensor during training. Since the model is too large, it would be nice to use ZeRO Stage 3 to shard the model weights, but it seems it is not possible since the scalar weight cannot be sharded across multiple GPUs.
How should one train a model with scalar weight in this case?
One way I thought about was having a vector of size (1 x # GPUs), then mean-reducing the vector before using it. If there is better or clever tricks, please let me know!
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
-
I have a problem where I need a scalar weight tensor during training. Since the model is too large, it would be nice to use ZeRO Stage 3 to shard the model weights, but it seems it is not possible since the scalar weight cannot be sharded across multiple GPUs.
How should one train a model with scalar weight in this case?
One way I thought about was having a vector of size (1 x # GPUs), then mean-reducing the vector before using it. If there is better or clever tricks, please let me know!
Beta Was this translation helpful? Give feedback.
All reactions