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 carefully checked the sample code and there seems to be no obvious error that would cause process 2 to transfer data to GPU 0.
So:
Why does process 2 allocate memory of GPU 0?
Is this part of the data involved in the calculation? I think if this part of the data is involved in the calculation when the number of processes becomes large, it will cause GPU 0 to be seriously overloaded?
Is there any way to avoid it?
Thanks in advance to partners in the PyTorch community for their hard work.