Cuda environment in docker

[LaurenceQ]

It is stated that Purdue Anvil provides 128c/256t AMD EPYC-7763 CPUs with NVIDIA A100 GPUs with CUDA 12.6 and Singularity, but there is no cuda environment in the docker. Will we need to install cuda by ourselves in the setup.sh in the submission?

[UDXS]

Hello @LaurenceQ

What you will need to install and setup in the setup.sh is the CUDA Toolkit. If you are using a library like Pytorch, then this is typically installed as a pip wheel along with the library. Otherwise, please see here: https://docs.nvidia.com/cuda/cuda-installation-guide-linux/

CUDA core libraries/drivers will be loaded by the singularity -nv flag. The NVIDIA Container Toolkit works in the same way (which you may test your setup.sh with the provided Docker).

Thanks,
Davit