I noticed when you install pytorch via pip, cuda & cudnn dependencies are also inclued in the pytorch binaries. I think it’s great, but when I compile pytorch from source, the binaries I got still need to link to local cuda libraries. So my question is how can I bundle cuda & cudnn into pytorch binaries when compiling from source? @rasbt @ptrblck
You could statically link all libraries and use the pytorch/builder
repository to see how the conda binaries and pip wheels are built.