The DGX A100 default OS (Ubuntu 18.04 based) install does not include CUDA development packages installed by default. From apt-cache search I can see there is a cuda-libraries-dev-11-0 package in the /var/lib/apt/lists/international.download.nvidia.com_dgx_repos_bionic_dists_bionic-4.99_multiverse_binary-amd64_Packages repository.
However, this repository is missing other packages like TensorRT needed to build tensorflow from source. So my plan is to install the
http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64/nvidia-machine-learning-repo-ubuntu1804_1.0.0-1_amd64.deb
repository to pull in TensorRT (libnvinfer) among other libs. I see this will also upgrade the CUDNN libraries too over what is in the DGX specific repository.
Does anyone see any obvious problems doing this will cause?
I have been told in the past to “stick to running things via docker images from NGC” but for various reason this is a bad fit for our development model and that is not a viable option 100% of the time.