DataLoader cost 273MiB gpu memory on GPU-0 for each worker

When I use DataLoader with multiple num-workers , I found that though i run my program on GPU1~GPU7, each worker will cost 273MB gpu memory on GPU0. Anyone knows how to fix this issue? In this case, I cannot use my GPU0 card, since them memory are used for data loading.

it’s a opencv gpu version problem. You can link to a cpu version opencv to fix that.

The current opencv is built with WITH_CUDA=OFF. Is it enough?

Thank you so much! Finally I solved this problem with -DWITH_CUDA=OFF -DWITH_OPENCL=OFF