Using GPU inside docker container - CUDA Version: N/A and torch.cuda.is_available returns False

Question

I'm trying to use GPU from inside my docker container. I'm using docker with version 19.03 on Ubuntu 18.04.

Outside the docker container if I run nvidia-smi I get the below output.

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.51.05    Driver Version: 450.51.05    CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            On   | 00000000:00:1E.0 Off |                    0 |
| N/A   30C    P8     9W /  70W |      0MiB / 15109MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

If I run the samething inside the container created from nvidia/cuda docker image, I get the same output as above and everything is running smoothly. torch.cuda.is_available() returns True.

But If I run the same nvidia-smi command inside any other docker container, it gives the following output where you can see that the CUDA Version is coming as N/A. Inside the containers torch.cuda.is_available() also returns False.

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.51.05    Driver Version: 450.51.05    CUDA Version: N/A      |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla T4            On   | 00000000:00:1E.0 Off |                    0 |
| N/A   30C    P8     9W /  70W |      0MiB / 15109MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

I have installed nvidia-container-toolkit using the following commands.

curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
curl -s -L https://nvidia.github.io/nvidia-docker/ubuntu18.04/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update
sudo apt-get install nvidia-container-toolkit
sudo systemctl restart docker

I started my containers using the following commands

sudo docker run --rm --gpus all nvidia/cuda nvidia-smi
sudo docker run -it --rm --gpus all ubuntu nvidia-smi

Answer 1

For anybody arriving here looking how to do it with docker compose, add to your service:

deploy:
  resources:
    reservations:
      devices:
      - driver: nvidia
        capabilities:
          - gpu
          - utility # nvidia-smi
          - compute # CUDA. Required to avoid "CUDA version: N/A"
          - video   # NVDEC/NVENC. For instance to use a hardware accelerated ffmpeg. Skip it if you don't need it

Doc: https://docs.docker.com/compose/gpu-support