The long version is that nvidia/cuda is now using CUDA 10.0, and it requires a recent driver.
With CUDA 10.0, this check is failing:
It's a logical OR between these two expressions:
Either you have a driver that supports CUDA 10.0. This is mentioned here: [login to view URL]
OR You have a 384 driver AND Tesla GPU.
Here, none of these conditions are true. And your error message is even weirder since you don't have the latest release of libnvidia-container, you should upgrade to 1.0.0 by upgrading to the latest package.
If you want to use CUDA 10.0, you need to upgrade your NVIDIA driver.
If you are fine with staying on CUDA 9.0, use image nvidia/cuda:9.0-base instead.
Above seems like the proposed solution need help in implementing it