gpu4singularity is deprecated for most uses (except maybe compiling against driver libs at bootstrap and some other weird stuff). If you are using Singularity >= 2.3, use the experimental option
--nv to grant your containers GPU support at runtime.
Singularity is a container platform that let’s you “swap” out your host operating system for one that you control.
Getting software within a container to recognize and use GPU hardware on the host can be tricky. gpu4singularity simplifies this process by automating the steps you must take to install GPU support within your container. Then you can use programs that rely on CUDA and cuDNN.
gpu4singularity does the following:
- Downloads an NVIDIA .run script for driver installation
- Extracts the contents (doesn’t actually install the driver)
- Moves all of the libraries to a central location (
- Creates a bunch of symbolic links
- Edits and exports the $PATH and $LD_LIBRARY_PATH to point to the correct libraries and binaries