Home

Awesome

GPU-Jupyter

<img src="https://raw.githubusercontent.com/iot-salzburg/gpu-jupyter/master/extra/logo_gupyter.png" alt="GPU-Jupyter" width=661/>

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

Github Workflow Docker Pulls Docker Stars GitHub stars GitHub forks GitHub pull-requests closed GitHub commits Maintenance

Jupyterlab Overview

Welcome to this project, which provides a GPU-capable environment based on NVIDIA's CUDA Docker image and the popular docker-stacks. Our toolstack enables GPU calculations in Jupyter notebooks, while the use of containers and versioned tags ensures the reproducibility of experiments. By utilizing version control for the source code, tagged data spaces, seeds for the random functions, and tagged environments as provided here, our solution empowers researchers to conduct highly complex big data experiments that are fully reproducible.

The images of this repository are available on Dockerhub. A big thank you to docker-stacks for creating and maintaining a robust Python, R, and Julia toolstack for Data Science.

Contents

  1. Quickstart
  2. Build Your image
  3. Tracing
  4. Configuration
  5. Issues and Contributing
  6. Support

Quickstart

  1. Ensure that you have access to a computer with an NVIDIA GPU. If you don’t have your own setup, you can try Saturn Cloud for a free GPU-powered Jupyter solution.

  2. Install Docker version 1.10.0+ and Docker Compose version 1.28.0+.

  3. Get access to your GPU via CUDA drivers within Docker containers. For this, follow the installation steps in this Medium article. You can confirm that you can access your GPU within Docker if the command below returns a result similar to this one:

    docker run --rm --gpus all nvidia/cuda:12.5.1-cudnn-runtime-ubuntu22.04 nvidia-smi

    Tue Nov 26 15:13:37 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 555.42.03              Driver Version: 555.85         CUDA Version: 12.5     |
|-----------------------------------------+------------------------+----------------------+
| 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  NVIDIA GeForce RTX 3060 ...    On  |   00000000:01:00.0 Off |                  N/A |
| N/A   43C    P8             12W /   60W |    4569MiB /   6144MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A       231      C   /python3.11                                 N/A      |
+-----------------------------------------------------------------------------------------+

    It is important to keep your installed CUDA version in mind when you pull images. Note that you can't run images based on nvidia/cuda:11.2 if you have only CUDA version 10.1 installed, use nvcc --version to get the correct cuda version. Additionally, a NVIDIA driver version of at least 520 is suggested, as the images are built and tested using this and later versions.

  4. Pull and run the GPU-Jupyter image. This may take some time as the whole environment for data science will be downloaded:

    cd your-working-directory
ll data  # this path will be mounted by default
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:v1.8_cuda-12.5_ubuntu-22.04

    This starts an instance of GPU-Jupyter with the tag v1.8_cuda-12.5_ubuntu-22.04 at http://localhost:8848 (port 8848). To log into Jupyterlab, you have to specify a token that you get from:

    docker exec -it [container-ID/name] jupyter server list
# [JupyterServerListApp] Currently running servers:
# [JupyterServerListApp] http://791003a731e1:8888/?token=5b96bb15be315ccb24643ea368a52cc0ba13657fbc29e409 :: /home/jovyan

    You can optionally set a password in http://localhost:8848/login or as described below (former default password gpu-jupyter). Additionally, data within the host's data directory is shared with the container.

    The following images of GPU-Jupyter are available on Dockerhub:

    • v1.8_cuda-12.5_ubuntu-22.04 (full image)
    • v1.8_cuda-12.5_ubuntu-22.04_python-only (only with a python interpreter and without Julia and R)
    • v1.8_cuda-12.5_ubuntu-22.04_slim (only with a python interpreter and without additional packages)
    • v1.7_cuda-12.3_ubuntu-22.04 (full image)
    • v1.7_cuda-12.3_ubuntu-22.04_python-only (only with a python interpreter and without Julia and R)
    • v1.7_cuda-12.3_ubuntu-22.04_slim (only with a python interpreter and without additional packages)
    • v1.6_cuda-12.0_ubuntu-22.04 (full image)
    • v1.6_cuda-12.0_ubuntu-22.04_python-only (only with a python interpreter and without Julia and R)
    • v1.6_cuda-12.0_ubuntu-22.04_slim (only with a python interpreter and without additional packages)
    • v1.6_cuda-11.8_ubuntu-22.04 (full image)
    • v1.6_cuda-11.8_ubuntu-22.04_python-only (only with a python interpreter and without Julia and R)
    • v1.6_cuda-11.8_ubuntu-22.04_slim (only with a python interpreter and without additional packages)
    <details> <summary><font color=blue> Older images</font></summary>
    • v1.5_cuda-12.0_ubuntu-22.04 (full image)
    • v1.5_cuda-12.0_ubuntu-22.04_python-only (only with a python interpreter and without Julia and R)
    • v1.5_cuda-12.0_ubuntu-22.04_slim (only with a python interpreter and without additional packages)
    • v1.5_cuda-11.8_ubuntu-22.04 (full image)
    • v1.5_cuda-11.8_ubuntu-22.04_python-only (only with a python interpreter and without Julia and R)
    • v1.5_cuda-11.8_ubuntu-22.04_slim (only with a python interpreter and without additional packages)
    • v1.5_cuda-11.6_ubuntu-20.04 (full image)
    • v1.5_cuda-11.6_ubuntu-20.04_python-only (only with a python interpreter and without Julia and R)
    • v1.5_cuda-11.6_ubuntu-20.04_slim (only with a python interpreter and without additional packages)
    • v1.4_cuda-11.6_ubuntu-20.04 (full image)
    • v1.4_cuda-11.6_ubuntu-20.04_python-only (only with a python interpreter and without Julia and R)
    • v1.4_cuda-11.6_ubuntu-20.04_slim (only with a python interpreter and without additional packages)
    • v1.4_cuda-11.2_ubuntu-20.04 (full image)
    • v1.4_cuda-11.2_ubuntu-20.04_python-only (only with a python interpreter and without Julia and R)
    • v1.4_cuda-11.2_ubuntu-20.04_slim (only with a python interpreter and without additional packages)
    • v1.4_cuda-11.0_ubuntu-20.04 (full image)
    • v1.4_cuda-11.0_ubuntu-20.04_python-only (only with a python interpreter and without Julia and R)
    • v1.4_cuda-11.0_ubuntu-20.04_slim (only with a python interpreter and without additional packages)
    • v1.4_cuda-11.0_ubuntu-18.04 (full image)
    • v1.4_cuda-11.0_ubuntu-18.04_python-only (only with a python interpreter and without Julia and R)
    • v1.4_cuda-11.0_ubuntu-18.04_slim (only with a python interpreter and without additional packages)
    • v1.4_cuda-10.1_ubuntu-18.04 (full image)
    • v1.4_cuda-10.1_ubuntu-18.04_python-only (only with a python interpreter and without Julia and R)
    • v1.4_cuda-10.1_ubuntu-18.04_slim (only with a python interpreter and without additional packages)
    </details>

    The version, e.g. v1.8, declares the version of the generator setup. The Cuda version, e.g. cuda-12.5, must match the CUDA driver version and be supported by the GPU libraries. These and older versions of GPU-Jupyter are listed on Dockerhub. In case you are using another version or the GPU libraries don't work on your hardware, please try to build the image on your own as described below. Note that the images built for Ubuntu 20.04 LTS work also on Ubuntu 22.04 LTS.

Within the Jupyterlab UI, ensure you can access your GPU by opening a new Terminal window and running nvidia-smi. Here, you can also install additional packages on top of the built image. To help you get started with using the GPU, the repository includes some sample code located in extra/Getting_Started. If you're new to JupyterLab or want to learn more about its features, we recommend checking out this tutorial.

Build Your Image

Building a custom Docker image is the recommended option if you have a different GPU architecture or if you want to customize the pre-installed packages. The Dockerfiles in custom/ can be modified to achieve this. To use a custom base image, modify custom/header.Dockerfile. To install specific GPU-related libraries, modify custom/gpulibs.Dockerfile, and to add specific libraries, append them to custom/usefulpackages.Dockerfile. Moreover, this offers the option for a static token or password which does not change with a container's restart.

After making the necessary modifications, regenerate the Dockerfile in /.build. Once you have confirmed that your GPU is accessible within Docker containers by running docker run --gpus all nvidia/cuda:12.5.1-cudnn-runtime-ubuntu22.04 nvidia-sm and seeing the GPU statistics, you can generate, build, and run the Docker image. The following commands will start GPU-Jupyter on localhost:8848 with the default password gpu-jupyter.

git clone https://github.com/iot-salzburg/gpu-jupyter.git
cd gpu-jupyter
git branch  # Check for extisting branches
git checkout v1.8_cuda-12.5_ubuntu-22.04  # select or create a new version
# generate the Dockerfile with python and without Julia and R (see options: --help)
./generate-Dockerfile.sh --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 command starts a container with GPU support and a shared local data volume /data, along with other necessary configurations, such as root permissions to install packages within the container. For more configuration options, see Configuration of the Dockerfile-Generation below.

Start via Docker Compose

To start GPU-Jupyter using docker-compose.yml, run the following command:

docker-compose up --build -d  # build and run in detached mode
docker-compose ps  # check if was successful
docker-compose logs -f  # view the logs
docker-compose down  # stop the container

This step requires a docker-compose version of at least 1.28.0, as the Dockerfile requests GPU resources (see this changelog). To update docker-compose, this discussion may be useful.

Tracing

With these commands we can investigate the container:

docker ps  # use the flat '-a' to view all
docker stats
docker logs [service-name | UID] -f  # view the logs
docker exec -it [service-name | UID] bash  # open bash in the container

To stop the local deployment, run:

docker rm -f [service-name | UID]  # or

Configuration

Configuration of the Dockerfile-Generation

To generate a Dockerfile for GPU-Jupyter, you can use the script generate-Dockerfile.sh. It generates a Dockerfile in the .build/ directory, which is overwritten with each new generation and entries within this directory should not be modified manually.

The script has the following parameters:

Note that only one of the parameters --slim, --python-only, and --no-useful-packages can be used at the same time:

Custom Installations

If you need to install custom packages within the container, you can modify the custom/usefulpackages.Dockerfile file or do it directly within the container. Keep in mind that every time a Dockerfile is generated, the file .build/Dockerfile is overwritten, so it's best to append custom installations in custom/usefulpackages.Dockerfile or generate-Dockerfile.sh.

Some useful packages are suggested in the Extension docs and in this blog article from neptune.ai. If you notice that an important package is missing in the default stack, please let us know so we can update it.

Authorization

Set a Static Token

Jupyter by default regenerates a new token on each new start. GPU-Jupyter provides the environment variable JUPYTER_TOKEN to set a customized static token. This option is practicable if the host machine is periodically restartet. It is suggested to use a long token such as a UUID:

export JUPYTER_TOKEN=$(uuidgen)
echo $JUPYTER_TOKEN

The static token can be requested using docker exec:

docker exec -it gpu-jupyter_1 jupyter server list

Set a custom Password

There are two ways to set a password for GPU-Jupyter:

  1. Go to the login page (http://192.168.48.48:8848/login) when logged out and setup a Password in the corresponding field.

  2. Use the --password or --pw option in the generate-Dockerfile.sh script to specify your desired password, like so:

    bash generate-Dockerfile.sh --password [your_password]

    This will update automatically the salted hashed token in the .build/jupyter_notebook_config.json file. Note that the specified password may be visible in your account's bash history.

Adaptions for using Tensorboard

Both TensorFlow and PyTorch support tensorboard. This packages is already installed in the GPU-packages and can be used with these settings:

  1. Forward the port in the docker command using -p 6006:6006 (only for usage outside of Juypterlab).
  2. Starting tensorboad with port binding within a container or Jupyterlab UI. Make sure the parameter --bind_all is set.
docker exec -it [container-name/ID] bash
root@749eb1a06d60:~# tensorboard --logdir mylogdir --bind_all

%tensorboard --logdir logs/[logdir] --bind_all
  1. Writing the states and results in the tensorboard log-dir, as described in the tutorials for TensorFlow and PyTorch or in the Getting Started section data/Getting_Started. If the port is exposed, tensorboard can be accessed in the browser on localhost:6006.

Updates

Update CUDA to another version

The GPU-libraries such as PyTorch and Tensorflow in custom/Docker.gpulibs must support the CUDA version and NVIDIA drivers on the host machine. Check out the compatibility lists for PyTorch and Tensorflow or search online for the explicit versions. In this setup, the NVIDIA Driver has version 530.30.02 and CUDA version 11.6.2 is used, which is compatible with Tensorflow 2.10 and PyTorch 1.12.

The host's CUDA version must be equal to or higher than that used by the container (set within custom/header.Dockerfile). 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 (find all tags here) in custom/header.Dockerfile to, e.g.: the line:

FROM nvidia/cuda:X.Y-base-ubuntu20.04

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

./generate-Dockerfile.sh --slim  # generate the Dockerfile with only a python interpreter, --python-only is default
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

Upgrade / Downgrade the Docker-Stacks

The docker-stacks is used as a submodule within .build/docker-stacks. Per default, the head of the commit is reset to a commit on which gpu-jupyter runs stable. To set the docker-stacks to a specific version generate the Dockerfile with a specific docker-stacks commit, run:

./generate-Dockerfile.sh --commit c1c32938438151c7e2a22b5aa338caba2ec01da2

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

./generate-Dockerfile.sh --commit latest

A new build can last some time and may consume a lot of data traffic. Note, that untested versions often result in a version conflict, as some files have to be adapted. Here are some examples of solutions:

Configure a 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 docker-compose.yml:

services:
  data-source-service:
  ...
      networks:
      - default
      - datastack
  ...
networks:
  datastack:
    driver: overlay
    attachable: true

In this example, the docker network has the name datastack as defined within the docker-compose.yml file and is configured to be attachable.

Issues and Contributing

Frequent Issues:

Contribution

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

Support

For commercial support, please contact b-data by email: support@b-data.ch.

b-data maintains its own CUDA-enabled JupyterLab docker stacks and tailors them to your needs, e.g.

Additionally, the JupyterHub setup can be customised to allow

and much more.