Nvidia-docker: Unsupported CUDA version: driver 0.0 < image 8.0 on Ubuntu 16.04

[Quintasan]

Hi, I'm trying to get anything running using nvidia-docker on my ThinkPad T430 (NVIDIA Corporation GF108M [NVS 5400M] (rev a1)).

Here's is what I did (this a clean install, nothing more):

1. Install Ubuntu 16.04
2. Install nvidia-375 from graphic-drivers ppa
3. Install nvidia-modprobe
4. Install nvidia-docker from deb package

nvidia-smi

Tue Nov 29 17:51:54 2016       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 375.20                 Driver Version: 375.20                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  NVS 5400M           Off  | 0000:01:00.0     N/A |                  N/A |
| N/A   55C    P8    N/A /  N/A |    232MiB /   993MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0                  Not Supported                                         |
+-----------------------------------------------------------------------------+

nvidia-docker run --rm nvidia/cuda nvidia-smi

nvidia-docker | 2016/11/29 17:50:59 Error: unsupported CUDA version: driver 0.0 < image 8.0

It appears that nvidia-docker.service is not running. When I tried starting it with sudo systemctl start nvidia-docker.service it failed with the following message (from journalctl -xe):

lis 29 17:51:46 tyria nvidia-docker-plugin[3884]: /usr/bin/nvidia-docker-plugin | 2016/11/29 17:51:46 Loading NVIDIA unified memory
lis 29 17:51:46 tyria nvidia-docker-plugin[3884]: /usr/bin/nvidia-docker-plugin | 2016/11/29 17:51:46 Loading NVIDIA management library
lis 29 17:51:46 tyria nvidia-docker-plugin[3884]: /usr/bin/nvidia-docker-plugin | 2016/11/29 17:51:46 Discovering GPU devices
lis 29 17:51:46 tyria nvidia-docker-plugin[3884]: /usr/bin/nvidia-docker-plugin | 2016/11/29 17:51:46 Error: cuda: CUDA driver version is insufficient for CUDA runtime version

What am I doing wrong?

[flx42]

First of all, this is GPU architecture is now marked as deprecated in CUDA 8.0. In the next CUDA version, you won't be able to compile CUDA code for your architecture.
It also means NVML (one of the library we use) might not work very well with this architecture.

But anyway, you shouldn't install the drivers with the ppa, you should install the official drivers. You should try that first.

[Quintasan]

@flx42 Any particular reason for not using that PPA? The version does match. I'll report back when I'm done with reinstalling them.

[Quintasan]

Welp, after installing drivers from the binary I managed to get the nvidia-docker running but my X server is no longer functioning. As it stands I have wasted over 3 days trying to get this working only to end up with with a subpar solution.

[flx42]

You can try nvidia-xconfig.

[flx42]

@Quintasan I stumbled into the same problem with the PPA actually, you just need to install an extra package: libcuda1-375

[3XX0]

Generally we don't recommend installing the driver from third party PPAs.
The official way of installing the CUDA driver is described in #258