Using GPU from a docker container?

I'm searching for a way to use the GPU from inside a docker container.

The container will execute arbitrary code so i don't want to use the privileged mode.

Any tips?

From previous research i understood that run -v and/or LXC cgroup was the way to go but i'm not sure how to pull that off exactly

Regan's answer is great, but it's a bit out of date, since the correct way to do this is avoid the lxc execution context as Docker has dropped LXC as the default execution context as of docker 0.9.

Instead it's better to tell docker about the nvidia devices via the --device flag, and just use the native execution context rather than lxc.

Environment

These instructions were tested on the following environment:

• Ubuntu 14.04
• CUDA 6.5
• AWS GPU instance.

Install nvidia driver and cuda on your host

See CUDA 6.5 on AWS GPU Instance Running Ubuntu 14.04 to get your host machine setup.

Install Docker

$ sudo apt-key adv --keyserver hkp://keyserver.ubuntu.com:80 --recv-keys 36A1D7869245C8950F966E92D8576A8BA88D21E9
$ sudo sh -c "echo deb https://get.docker.com/ubuntu docker main > /etc/apt/sources.list.d/docker.list"
$ sudo apt-get update && sudo apt-get install lxc-docker

Find your nvidia devices

ls -la /dev | grep nvidia

crw-rw-rw-  1 root root    195,   0 Oct 25 19:37 nvidia0 
crw-rw-rw-  1 root root    195, 255 Oct 25 19:37 nvidiactl
crw-rw-rw-  1 root root    251,   0 Oct 25 19:37 nvidia-uvm

Run Docker container with nvidia driver pre-installed

I've created a docker image that has the cuda drivers pre-installed. The dockerfile is available on dockerhub if you want to know how this image was built.

You'll want to customize this command to match your nvidia devices. Here's what worked for me:

 $ sudo docker run -ti --device /dev/nvidia0:/dev/nvidia0 --device /dev/nvidiactl:/dev/nvidiactl --device /dev/nvidia-uvm:/dev/nvidia-uvm tleyden5iwx/ubuntu-cuda /bin/bash

Verify CUDA is correctly installed

This should be run from inside the docker container you just launched.

Install CUDA samples:

$ cd /opt/nvidia_installers
$ ./cuda-samples-linux-6.5.14-18745345.run -noprompt -cudaprefix=/usr/local/cuda-6.5/

Build deviceQuery sample:

$ cd /usr/local/cuda/samples/1_Utilities/deviceQuery
$ make
$ ./deviceQuery   

If everything worked, you should see the following output:

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 6.5, CUDA Runtime Version = 6.5, NumDevs =    1, Device0 = GRID K520
Result = PASS

Writing an updated answer since most of the already present answers are obsolete as of now.

Versions earlier than Docker 19.03 used to require nvidia-docker2 and the --runtime=nvidia flag.

Since Docker 19.03, you need to install nvidia-container-toolkit package and then use the --gpus all flag.

So, here are the basics,

Package Installation

Install the nvidia-container-toolkit package as per official documentation at Github.

For Redhat based OSes, execute the following set of commands:

$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
$ curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.repo | sudo tee /etc/yum.repos.d/nvidia-docker.repo

$ sudo yum install -y nvidia-container-toolkit
$ sudo systemctl restart docker

For Debian based OSes, execute the following set of commands:

# Add the package repositories
$ distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
$ curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
$ curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list

$ sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit
$ sudo systemctl restart docker

Running the docker with GPU support

docker run --name my_all_gpu_container --gpus all -t nvidia/cuda

Please note, the flag --gpus all is used to assign all available gpus to the docker container.

To assign specific gpu to the docker container (in case of multiple GPUs available in your machine)

docker run --name my_first_gpu_container --gpus device=0 nvidia/cuda

Or

docker run --name my_first_gpu_container --gpus '"device=0"' nvidia/cuda

Ok i finally managed to do it without using the --privileged mode.

I'm running on ubuntu server 14.04 and i'm using the latest cuda (6.0.37 for linux 13.04 64 bits).

---

#Preparation Install nvidia driver and cuda on your host. (it can be a little tricky so i will suggest you follow this guide https://askubuntu.com/questions/451672/installing-and-testing-cuda-in-ubuntu-14-04)

ATTENTION : It's really important that you keep the files you used for the host cuda installation

---

#Get the Docker Daemon to run using lxc We need to run docker daemon using lxc driver to be able to modify the configuration and give the container access to the device.

One time utilization :

sudo service docker stop
sudo docker -d -e lxc

Permanent configuration Modify your docker configuration file located in /etc/default/docker Change the line DOCKER_OPTS by adding '-e lxc' Here is my line after modification

DOCKER_OPTS="--dns 8.8.8.8 --dns 8.8.4.4 -e lxc"

Then restart the daemon using

sudo service docker restart

How to check if the daemon effectively use lxc driver ?

docker info

The Execution Driver line should look like that :

Execution Driver: lxc-1.0.5

---

#Build your image with the NVIDIA and CUDA driver. Here is a basic Dockerfile to build a CUDA compatible image.

FROM ubuntu:14.04
MAINTAINER Regan <http://stackoverflow.com/questions/25185405/using-gpu-from-a-docker-container>

RUN apt-get update && apt-get install -y build-essential
RUN apt-get --purge remove -y nvidia*

ADD ./Downloads/nvidia_installers /tmp/nvidia                             > Get the install files you used to install CUDA and the NVIDIA drivers on your host
RUN /tmp/nvidia/NVIDIA-Linux-x86_64-331.62.run -s -N --no-kernel-module   > Install the driver.
RUN rm -rf /tmp/selfgz7                                                   > For some reason the driver installer left temp files when used during a docker build (i don't have any explanation why) and the CUDA installer will fail if there still there so we delete them.
RUN /tmp/nvidia/cuda-linux64-rel-6.0.37-18176142.run -noprompt            > CUDA driver installer.
RUN /tmp/nvidia/cuda-samples-linux-6.0.37-18176142.run -noprompt -cudaprefix=/usr/local/cuda-6.0   > CUDA samples comment if you don't want them.
RUN export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64         > Add CUDA library into your PATH
RUN touch /etc/ld.so.conf.d/cuda.conf                                     > Update the ld.so.conf.d directory
RUN rm -rf /temp/*  > Delete installer files.

---

#Run your image. First you need to identify your the major number associated with your device. Easiest way is to do the following command :

ls -la /dev | grep nvidia

If the result is blank, use launching one of the samples on the host should do the trick. The result should look like that ![enter image description here][1] [1]: http://i.stack.imgur.com/j8Jr3.png As you can see there is a set of 2 numbers between the group and the date. These 2 numbers are called major and minor numbers (wrote in that order) and design a device. We will just use the major numbers for convenience.

Why do we activated lxc driver? To use the lxc conf option that allow us to permit our container to access those devices. The option is : (i recommend using * for the minor number cause it reduce the length of the run command)

> --lxc-conf='lxc.cgroup.devices.allow = c [major number]:[minor number or *] rwm'

So if i want to launch a container (Supposing your image name is cuda).

docker run -ti --lxc-conf='lxc.cgroup.devices.allow = c 195:* rwm' --lxc-conf='lxc.cgroup.devices.allow = c 243:* rwm' cuda

We just released an experimental GitHub repository which should ease the process of using NVIDIA GPUs inside Docker containers.

Recent enhancements by NVIDIA have produced a much more robust way to do this.

Essentially they have found a way to avoid the need to install the CUDA/GPU driver inside the containers and have it match the host kernel module.

Instead, drivers are on the host and the containers don't need them. It requires a modified docker-cli right now.

This is great, because now containers are much more portable.

A quick test on Ubuntu:

# Install nvidia-docker and nvidia-docker-plugin
wget -P /tmp https://github.com/NVIDIA/nvidia-docker/releases/download/v1.0.1/nvidia-docker_1.0.1-1_amd64.deb
sudo dpkg -i /tmp/nvidia-docker*.deb && rm /tmp/nvidia-docker*.deb

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

For more details see: GPU-Enabled Docker Container and: https://github.com/NVIDIA/nvidia-docker

Updated for cuda-8.0 on ubuntu 16.04

• Install docker https://www.digitalocean.com/community/tutorials/how-to-install-and-use-docker-on-ubuntu-16-04

• Build the following image that includes the nvidia drivers and the cuda toolkit

Dockerfile

FROM ubuntu:16.04
MAINTAINER Jonathan Kosgei <[email protected]>

# A docker container with the Nvidia kernel module and CUDA drivers installed

ENV CUDA_RUN https://developer.nvidia.com/compute/cuda/8.0/prod/local_installers/cuda_8.0.44_linux-run

RUN apt-get update && apt-get install -q -y \
  wget \
  module-init-tools \
  build-essential 

RUN cd /opt && \
  wget $CUDA_RUN && \
  chmod +x cuda_8.0.44_linux-run && \
  mkdir nvidia_installers && \
  ./cuda_8.0.44_linux-run -extract=`pwd`/nvidia_installers && \
  cd nvidia_installers && \
  ./NVIDIA-Linux-x86_64-367.48.run -s -N --no-kernel-module

RUN cd /opt/nvidia_installers && \
  ./cuda-linux64-rel-8.0.44-21122537.run -noprompt

# Ensure the CUDA libs and binaries are in the correct environment variables
ENV LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-8.0/lib64
ENV PATH=$PATH:/usr/local/cuda-8.0/bin

RUN cd /opt/nvidia_installers &&\
    ./cuda-samples-linux-8.0.44-21122537.run -noprompt -cudaprefix=/usr/local/cuda-8.0 &&\
    cd /usr/local/cuda/samples/1_Utilities/deviceQuery &&\ 
    make

WORKDIR /usr/local/cuda/samples/1_Utilities/deviceQuery

3. Run your container

sudo docker run -ti --device /dev/nvidia0:/dev/nvidia0 --device /dev/nvidiactl:/dev/nvidiactl --device /dev/nvidia-uvm:/dev/nvidia-uvm <built-image> ./deviceQuery

You should see output similar to:

deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 8.0, CUDA Runtime Version = 8.0, NumDevs = 1, Device0 = GRID K520 Result = PASS

Goal:

My goal was to make a CUDA enabled docker image without using nvidia/cuda as base image. Because I have some custom jupyter image, and I want to base from that.

Prerequisite:

The host machine had nvidia driver, CUDA toolkit, and nvidia-container-toolkit already installed. Please refer to the official docs, and to Rohit's answer.

Test that nvidia driver and CUDA toolkit is installed correctly with: nvidia-smi on the host machine, which should display correct "Driver Version" and "CUDA Version" and shows GPUs info.

Test that nvidia-container-toolkit is installed correctly with: docker run --rm --gpus all nvidia/cuda:latest nvidia-smi

Dockerfile

I found what I assume to be the official Dockerfile for nvidia/cuda here I "flattened" it, appended the contents to my Dockerfile and tested it to be working nicely:

FROM sidazhou/scipy-notebook:latest
# FROM ubuntu:18.04 

###########################################################################
# See https://gitlab.com/nvidia/container-images/cuda/-/blob/master/dist/10.1/ubuntu18.04-x86_64/base/Dockerfile
# See https://sarus.readthedocs.io/en/stable/user/custom-cuda-images.html
###########################################################################
USER root

###########################################################################
# base
RUN apt-get update && apt-get install -y --no-install-recommends \
    gnupg2 curl ca-certificates && \
    curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub | apt-key add - && \
    echo "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64 /" > /etc/apt/sources.list.d/cuda.list && \
    echo "deb https://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1804/x86_64 /" > /etc/apt/sources.list.d/nvidia-ml.list && \
    apt-get purge --autoremove -y curl \
    && rm -rf /var/lib/apt/lists/*

ENV CUDA_VERSION 10.1.243
ENV CUDA_PKG_VERSION 10-1=$CUDA_VERSION-1

# For libraries in the cuda-compat-* package: https://docs.nvidia.com/cuda/eula/index.html#attachment-a
RUN apt-get update && apt-get install -y --no-install-recommends \
    cuda-cudart-$CUDA_PKG_VERSION \
    cuda-compat-10-1 \
    && ln -s cuda-10.1 /usr/local/cuda && \
    rm -rf /var/lib/apt/lists/*

# Required for nvidia-docker v1
RUN echo "/usr/local/nvidia/lib" >> /etc/ld.so.conf.d/nvidia.conf && \
    echo "/usr/local/nvidia/lib64" >> /etc/ld.so.conf.d/nvidia.conf

ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:${PATH}
ENV LD_LIBRARY_PATH /usr/local/nvidia/lib:/usr/local/nvidia/lib64


###########################################################################
#runtime next
ENV NCCL_VERSION 2.7.8

RUN apt-get update && apt-get install -y --no-install-recommends \
    cuda-libraries-$CUDA_PKG_VERSION \
    cuda-npp-$CUDA_PKG_VERSION \
    cuda-nvtx-$CUDA_PKG_VERSION \
    libcublas10=10.2.1.243-1 \
    libnccl2=$NCCL_VERSION-1+cuda10.1 \
    && apt-mark hold libnccl2 \
    && rm -rf /var/lib/apt/lists/*

# apt from auto upgrading the cublas package. See https://gitlab.com/nvidia/container-images/cuda/-/issues/88
RUN apt-mark hold libcublas10


###########################################################################
#cudnn7 (not cudnn8) next

ENV CUDNN_VERSION 7.6.5.32

RUN apt-get update && apt-get install -y --no-install-recommends \
    libcudnn7=$CUDNN_VERSION-1+cuda10.1 \
    && apt-mark hold libcudnn7 && \
    rm -rf /var/lib/apt/lists/*


ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES all
ENV NVIDIA_REQUIRE_CUDA "cuda>=10.1"


###########################################################################
#docker build -t sidazhou/scipy-notebook-gpu:latest .

#docker run -itd -gpus all\
#  -p 8888:8888 \
#  -p 6006:6006 \
#  --user root \
#  -e NB_UID=$(id -u) \
#  -e NB_GID=$(id -g) \
#  -e GRANT_SUDO=yes \
#  -v ~/workspace:/home/jovyan/work \
#  --name sidazhou-jupyter-gpu \
#  sidazhou/scipy-notebook-gpu:latest

#docker exec sidazhou-jupyter-gpu python -c "import tensorflow as tf; print(tf.config.experimental.list_physical_devices('GPU'))"

To use GPU from docker container, instead of using native Docker, use Nvidia-docker. To install Nvidia docker use 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/ubuntu16.04/amd64/nvidia-
docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update
sudo apt-get install -y nvidia-docker
sudo pkill -SIGHUP dockerd # Restart Docker Engine
sudo nvidia-docker run --rm nvidia/cuda nvidia-smi # finally run nvidia-smi in the same container

Use x11docker by mviereck:

https://github.com/mviereck/x11docker#hardware-acceleration says

> Hardware acceleration

> Hardware acceleration for OpenGL is possible with option -g, --gpu.

> This will work out of the box in most cases with open source drivers on host. Otherwise have a look at wiki: feature dependencies. Closed source NVIDIA drivers need some setup and support less x11docker X server options.

This script is really convenient as it handles all the configuration and setup. Running a docker image on X with gpu is as simple as

x11docker --gpu imagename

I would not recommend installing CUDA/cuDNN on the host if you can use docker. Since at least CUDA 8 it has been possible to "stand on the shoulders of giants" and use nvidia/cuda base images maintained by NVIDIA in their Docker Hub repo. Go for the newest and biggest one (with cuDNN if doing deep learning) if unsure which version to choose.

A starter CUDA container:

mkdir ~/cuda11
cd ~/cuda11

echo "FROM nvidia/cuda:11.0-cudnn8-devel-ubuntu18.04" > Dockerfile
echo "CMD [\"/bin/bash\"]" >> Dockerfile

docker build --tag mirekphd/cuda11 .

docker run --rm -it --gpus 1 mirekphd/cuda11 nvidia-smi

Sample output:

(if nvidia-smi is not found in the container, do not try install it there - it was already installed on thehost with NVIDIA GPU driver and should be made available from the host to the container system if docker has access to the GPU(s)):

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.57       Driver Version: 450.57       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  GeForce GTX 108...  Off  | 00000000:01:00.0  On |                  N/A |
|  0%   50C    P8    17W / 280W |    409MiB / 11177MiB |      7%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

Prerequisites

1. Appropriate NVIDIA driver with the latest CUDA version support to be installed first on the host (download it from NVIDIA Driver Downloads and then mv driver-file.run driver-file.sh && chmod +x driver-file.sh && ./driver-file.sh). These are have been forward-compatible since CUDA 10.1.

2. GPU access enabled in docker by installing sudo apt get update && sudo apt get install nvidia-container-toolkit (and then restarting docker daemon using sudo systemctl restart docker).