Extra GPU memory usage when building from master

1

Hi all. I recently build pytorch from master, and it cost more GPU memory compared to the released 0.2.0.
I test GPU memory usage using the official imagenet example(Titan Xp, batchsize32). The numbers are as following:

Arch master 0.2.0
resnet50 5215MB 4169MB
resnet101 7547MB 5711MB
resnet152 10217MB 7495MB

As you can see — quite a lot. Also I have tried the 0.3.0 branch, where the GPU memory usage is almost the same with 0.2.0.
Anyone have the same issue?

2

Yeah, there is a bug on GPU memory management with views on master. So please use 0.3 branch for now :slight_smile:
This issue is relevant: https://github.com/pytorch/pytorch/issues/3610

3

Get it, thank you!:wink: