Hello, I deploy in an Openstack cloud a kubernetes solution with Magnum (container orchestration service for Openstack) and GPU aware nodes/containers for machine learning. Because cuda 9.0 (tensorflow doesn't support 9.1) is not available for other fedora releases than 25, and magnum doesn't support fedora 27, I naturally chose a fedora 25 atomic image for my cluster nodes. But I had a kernel panic with the nouveau driver on the fedora 25 guests, tried to disabling it, and never succeed. I tried again with a Fedora 27 atomic image and was able to disable nouveau without any problem. What is wrong with blacklisting on fedora 25 atomic? As long as Openstack magnum and Nvidia cuda doesn't provide fedora 27 support I can't just switch to fedora 27.
Hi @olivierdj. Wow! That really does sound like dependency hell.
So tensorflow doesn't support cuda > 9.0 which doesn't support Fedora > 25? Does newer tensorflow support cuda > 9.0 ?
Ultimately fedora 25 is End-Of-Life so there's not much use in trying to track down a bug there. Could you possibly try to install cuda 9.0 on F26 atomic (which is what openstack magnum is on now) and see if it happens to work?
Metadata Update from @dustymabe: - Issue tagged with: bug, host
closing this. @olivierdj please re-open if you want to continue the discussion
Metadata Update from @dustymabe: - Issue close_status updated to: Invalid - Issue status updated to: Closed (was: Open)
Log in to comment on this ticket.