Low GPU utility

[TianrongChen]

Hi Song Yang! Thank you for your code!
I am running your Network in a local (single GPU: RTX TITAN) machine, but the GPU utility is only around 30% during sampling. Did you notice the same issue?
Best Wishes,
Tianrong

[yang-song]

I didn't notice this issue before. Sometimes the CPU can also be a bottleneck. Maybe it will be better when using a more powerful CPU?

[TianrongChen]

Thank you for your reply! I figured it out XD
I think in my own code,
torch.backends.cudnn.deterministic = True
caused this problem.
Just turn this off, the problem will be solved.

[yang-song]

Great to know you figured it out! Yes you're right. Setting torch.backends.cudnn.deterministic = True ensures reproducibility but can greatly reduce the efficiency of GPUs.