Installing dcgm on a stateless node is untenable at the moment, because the dcgm package is a 1.5GB behemoth of which ~1GB are different versions of CUBLAS and ~240MB are different versions of CURAND. Appropiate versions of both libraries are usually present somewhere else in the system as they are pretty much essential for CUDA applications.
Please consider packaging those libraries separately and letting us point DCGM to our own location for those libraries.
Installing dcgm on a stateless node is untenable at the moment, because the dcgm package is a 1.5GB behemoth of which ~1GB are different versions of CUBLAS and ~240MB are different versions of CURAND. Appropiate versions of both libraries are usually present somewhere else in the system as they are pretty much essential for CUDA applications.
Please consider packaging those libraries separately and letting us point DCGM to our own location for those libraries.