I’m new to the forum, only here to reply to one post to share for people in my position.
I’m trying to get AI working on AWS p2 which is tesla K80 (compute capacity 3.7). This follows on from:
which is closed, advising the op to give up and use something else.
However, the python dependency can be ignored by installing a fake rpm providing ‘python(abi) = 3.6’
To do this, you just need to create an rpm with rpmbuild which does nothing except somewhere has
provides: python(abi) = 3.6
in the spec file
install, and the drivers from the rhel8 repo, nvidia-driver:470-dkms will work.
(However, my current problem is that, even if I install cuda and torch to support compute capability 3.7, they still require cublasGemmEx which is only supported from 5.0. My advice, don’t use p2. Apparently, this began at cuda8: