Gpus, nvidia-docker and rancher

We are beginning to deploy GPU enabled code and would like to tie this into our container deployment methods using rancher. Nvidia is distributing a thin-layer over docker called nvidia-docker which appears to smooth some of the hiccups between CUDA versions in containers and nvidia drivers residing on the physical hosts.

I am wondering if I can specify control how rancher launches containers on the my hosts and whether I can can set it up to use nvidia-docker instead of launching with docker?

I may going about this in a completely incorrect way as https://github.com/NVIDIA/nvidia-docker/wiki/NVIDIA-driver
implies you can do most without but with a lot poking around. Other suggestions welcome!

We are preparing for the same. the usage of CUDA is way forward for machine learning and other applications. Is this currently supported by Rancher? Any suggestion would be appreciated.

1 Like

Just to complete this thread. There is a more detailed discussion in https://github.com/rancher/rancher/issues/4883