Pytorch 1.6.0 seem to leak memory in conv2d

No idea, sorry. Check topics like this. Fragmentation may be an issue, if you’re stress testing.