I am working on a POC of running tensorflow container via Singularity 3.6. I have a requirement that the user may need to pick an older version of CUDA which are available in a customized location rather than the system installed one.
I read the 3.6 document for GPGPU support, the -nv flag does all magic automatically, I could not see any step where I can specific which CUDA to use. I have multiple CUDA versions on the system and they are summoned by module load command normally.
Any idea is welcomed.
Cheers