Need help with gpu-jupyter?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

168 Stars 64 Forks Apache License 2.0 150 Commits 5 Opened issues


Leverage the flexibility of Jupyterlab through the power of your NVIDIA GPU to run your code from Tensorflow and Pytorch in collaborative notebooks on the GPU.

Services available


Need anything else?

Contributors list


Leverage Jupyter Notebooks with the power of your NVIDIA GPU and perform GPU calculations using Tensorflow and Pytorch in collaborative notebooks.

Jupyterlab Overview

First of all, thanks to docker-stacks for creating and maintaining a robost Python, R and Julia toolstack for Data Analytics/Science applications. This project uses the NVIDIA CUDA image as the base image and installs their toolstack on top of it to enable GPU calculations in the Jupyter notebooks. The image of this repository is available on Dockerhub.


  1. Quickstart
  2. Build your own image
  3. Tracing
  4. Configuration
  5. Deployment
  6. Issues and Contributing


  1. A computer with an NVIDIA GPU is required.
  2. Install Docker version 1.10.0+ and Docker Compose version 1.6.0+.
  3. Get access to your GPU via CUDA drivers within Docker containers. You can be sure that you can access your GPU within Docker, if the command

    docker run --gpus all nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04 nvidia-smi
    returns a result similar to this one: ```bash Tue Jan 5 09:38:21 2021
    +-----------------------------------------------------------------------------+ | NVIDIA-SMI 450.80.02 Driver Version: 450.80.02 CUDA Version: 10.1 | |-------------------------------+----------------------+----------------------+ | 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 RTX 207... Off | 00000000:01:00.0 On | N/A | | 0% 40C P8 7W / 215W | 360MiB / 7974MiB | 1% Default | | | | N/A | +-------------------------------+----------------------+----------------------+

    +-----------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=============================================================================| +-----------------------------------------------------------------------------+ ``

    If you don't get an output similar than this one, follow the installation steps in this 
    [medium article](
    The CUDA toolkit is not required on the host system, as it will be 
    installed within the Docker containers using [NVIDIA-docker](
    It is also important to keep your installed CUDA version in mind, when you pull images. 
    **You can't run images based on
    if you have only CUDA version 10.1 installed.**
    Check your host's CUDA-version with
    nvcc --version` and update to at least the same version you want to pull.
  4. Pull and run the image. This can last some hours, as a whole data-science environment will be downloaded:

    cd your-working-directory 
    docker run --gpus all -d -it -p 8848:8888 -v $(pwd)/data:/home/jovyan/work -e GRANT_SUDO=yes -e JUPYTER_ENABLE_LAB=yes --user root cschranz/gpu-jupyter:v1.3_cuda-10.1_ubuntu-18.04_python-only
    This starts an instance with of GPU-Jupyter the tag
    at http://localhost:8848 (port
    ). The default password is
    ) which should be changed as described below. Furthermore, data within the host's
    directory is shared with the container. Other versions of GPU-Jupyter are available and listed on Dockerhub under Tags.

Within the Jupyterlab instance, you can check if you can access your GPU by opening a new terminal window and running

. In terminal windows, you can also install new packages for your own projects. Some example code can be found in the repository under
. If you want to learn more about Jupyterlab, check out this tutorial.

Build your own Image

First, it is necessary to generate the

, that is based on the NIVIDA base image and the docker-stacks. As soon as you have access to your GPU within Docker containers (make sure the command
docker run --gpus all nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04 nvidia-smi
shows your GPU statistics), you can generate the Dockerfile, build and run it. The following commands will start GPU-Jupyter on localhost:8848 with the default password
  git clone
  cd gpu-jupyter
  # generate a Dockerfile with python and without Julia and R
  ./ --python-only
  docker build -t gpu-jupyter .build/  # will take a while
  docker run --gpus all -d -it -p 8848:8888 -v $(pwd)/data:/home/jovyan/work -e GRANT_SUDO=yes -e JUPYTER_ENABLE_LAB=yes -e NB_UID="$(id -u)" -e NB_GID="$(id -g)" --user root --restart always --name gpu-jupyter_1 gpu-jupyter

This starts a container WITH GPU support, a shared local data volume


and some other configurations like root permissions which are necessary to install packages within the container. For more configurations, scroll down to Configuration of the Dockerfile-Generation.

Start via Docker Compose

The script
is a wrapper for a quick configuration of the underlying
  ./ -p 8848  # the default port is 8888


With these commands we can see if everything worked well:

docker ps
docker logs [service-name]  # or
bash  # a env-var safe wrapper for 'docker-compose logs -f'

In order to stop the local deployment, run:

docker ps
docker rm -f [service-name]  # or


Configuration of the Dockerfile-Generation

The script
generates a Dockerfile within the
directory. This implies that this Dockerfile is overwritten by each generation. The Dockerfile-generation script
has the following parameters (note that 2, 3 and 4 are exclusive):
  • -c|--commit
    : specify a commit or
    for the
    , the default commit is a working one.
  • -s|--slim
    : Generate a slim Dockerfile. As some installations are not needed by everyone, there is the possibility to skip some installations to reduce the size of the image. Here the
    is used instead of
    that comes with Julia and R. Moreover, none of the packages within
    is installed.
  • --python-only|--no-datascience-notebook
    : As the name suggests, the
    is not installed on top of the
    , but the packages within
  • --no-useful-packages
    : On top of the
    (Julia and R), the essential
    are installed, but not the packages within

Custom Installations

Custom packages can be installed within a container, or by modifying the file

. As
is overwritten each time a Dockerfile is generated, it is suggested to append custom installations either within
or in
If an essential package is missing in the default stack, please let us know!

Set Password

Please set a new password using

. Therefore, hash your password in the form (password)(salt) using a sha1 hash generator, e.g., the sha1 generator of The input with the default password
) is concatenated by an arbitrary salt
and is hashed to

Never give away your own unhashed password!

Then update the config file as shown below and restart the service.

  "NotebookApp": {
    "password": "sha1:3b4b6378355:642693b20f0a33bcad27b94293d0ed7db3408322"


Update CUDA to another version

The host's CUDA-version must be equal or higher than that of the container itself (in

). Check the host's version with
nvcc --version
and the version compatibilities for CUDA-dependent packages as Pytorch respectively Tensorflow previously. Then modify, if supported, the CUDA-version in
to, e.g.: the line:
FROM nvidia/cuda:11.1-base-ubuntu20.04

and in the

the line:

Then re-generate, re-build and run the updated image, as closer described above: Note that a change in the first line of the Dockerfile will re-build the whole image.

docker build -t gpu-jupyter .build/  # will take a while
docker run --gpus all -d -it -p 8848:8888 -v $pwd/data:/home/jovyan/work -e GRANT_SUDO=yes -e JUPYTER_ENABLE_LAB=yes --user root --restart always --name gpu-jupyter_1 gpu-jupyter 

Update Docker-Stack

The docker-stacks are used as a submodule within

. Per default, the head of the commit is reset to a commit on which
runs stable. To update the generated Dockerfile to a specific commit, run:
./ --commit c1c32938438151c7e2a22b5aa338caba2ec01da2

To update the generated Dockerfile to the latest commit, run:

./ --commit latest

A new build can last some time and may consume a lot of data traffic. Note, that the latest version may result in a version conflict! More info to submodules can be found in this tutorial.

Deployment in the Docker Swarm

A Jupyter instance often requires data from other services. If that data-source is containerized in Docker and sharing a port for communication shouldn't be allowed, e.g., for security reasons, then connecting the data-source with GPU-Jupyter within a Docker Swarm is a great option!

Set up Docker Swarm and Registry

This step requires a running Docker Swarm on a cluster or at least on this node. In order to register custom images in a local Docker Swarm cluster, a registry instance must be deployed in advance. Note that the we are using the port 5001, as many services use the default port 5000.

sudo docker service create --name registry --publish published=5001,target=5000 registry:2

This should output

. \

Afterwards, check if the registry service is available using

docker service ls

Configure the shared Docker network

Additionally, GPU-Jupyter is connected to the data-source via the same docker-network. Therefore, This network must be set to attachable in the source's

      - default
      - datastack
    driver: overlay
    attachable: true  

In this example, * the docker stack was deployed in Docker swarm with the name elk (

docker stack deploy ... elk
), * the docker network has the name datastack within the
file, * this network is configured to be attachable in the
file * and the docker network has the name elk_datastack, see the following output:
    sudo docker network ls
    # ...
    # [UID]        elk_datastack                   overlay             swarm
    # ...
The docker network name elk_datastack is used in the next step as a parameter.

Start GPU-Jupyter in Docker Swarm

Finally, GPU-Jupyter can be deployed in the Docker Swarm with the shared network, using:

./ -p [port] -n [docker-network] -r [registry-port]
# e.g. ./ -p 8848 -n elk_datastack -r 5001

where: * -p: port specifies the port on which the service will be available. * -n: docker-network is the name of the attachable network from the previous step, e.g., here it is elk_datastack. * -r: registry port is the port that is published by the registry service, default is


Now, gpu-jupyter will be accessible here on localhost:8848 with the default password

) and shares the network with the other data-source, i.e., all ports of the data-source will be accessible within GPU-Jupyter, even if they aren't routed it the source's

Check if everything works well using:

sudo docker service ps gpu_gpu-jupyter
docker service ps gpu_gpu-jupyter

In order to remove the service from the swarm, use:


Issues and Contributing

This project has the intention to create a robust image for CUDA-based GPU-applications, which is built on top of the docker-stacks. You are free to help to improve this project, by:

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.