PyTorch compiled with cuda and cudnn, so we don't need to build docker based on nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04.
ypw-rich opened this issue · 5 comments
PyTorch compiled with cuda and cudnn, so we don't need to build docker based on nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04.
cuda:10.0-base-ubuntu18.04 is only 115MB，but cuda:10.0-cudnn7-devel-ubuntu18.04 is 3.05GB.
cuda:10.0-base-ubuntu18.0 contains only the bare minimum (libcudart), while most DL frameworks depends on more CUDA packages?
Because PyTorch already compiled with cuda, it doesn't load cuda in the system, so we can add a if else to PyTorch to avoid overpacking.
@ypw-rich Thanks for the suggestion. Could you fix this and make a pull request? We will merge the code and trigger a rebuild for the related images.
This issue is stale because it has been open for 180 days with no activity.
This issue was closed because it has been inactive for 30 days since being marked as stale.