Same platform diff docker (runtime/devel) cant read GPU

all test in the same platform(centos, GPU driver is 416.xxx, )
in dockerhub
docker image pytorch/pytorch:1.8.1-cuda11.1-cudnn8-devel is ok, it can read gpu(torch.cuda.is_available() is true) but its too big, so i try pytorch/pytorch:1.8.1-cuda11.1-cudnn8-rutime cant read gpu ,why? where i can see the docker file? or how can i solve use samll docker image can read gpu question