py-torch by default enables mkldnn, which pulls in 5.4 GB of intel-oneapi-mkl. This disables the mkldnn functionality in py-torch.