How does one run the notebooks in the Docker container?

[cristobalmackenzie]

Hi

I use devcontainers for development all the time, but can't seem to figure out how to run the notebooks.

Can you point me in the right direction please? I'm currently running the notebooks on Colab but would rather just run them all from my local machine and keep the repo up to date.

[cristobalmackenzie]

jupyter lab --allow-root does the trick, is that the recommended way?

[rayed-therap]

This is not needed nor recommended. Please see my reply below.

[rasbt]

I use devcontainers for development all the time, but can't seem to figure out how to run the notebooks.

You could open and run them in Visual Studio code. Most people I know prefer it over Jupyter Lab. (Personally, while I use Visual Studio Code for basically all my coding, I do like Jupyter Lab UI a bit better though for notebooks; maybe muscle memory.)

jupyter lab --allow-root does the trick, is that the recommended way?

I wish I had a good answer for that. Personally, I am not a big devcontainer user and just use conda for managing virtual environments and packages on my computer (https://github.com/rasbt/LLMs-from-scratch/tree/main/appendix-A/01_optional-python-setup-preferences).

Maybe @rayedbw knows?

[d-kleine]

The good thing about using a Docker image is that you have also a common platform, which makes it easier for reproduction - it ensures consistency and reproducibility across different environments. When just working with a requirements.txt file, people most likely have different OSs, and sometimes those packages are broken because they will be developed on Linux and ported for instance to Windows. Also, some packages only work on Linux (I am just having this issue with triton right now for PyTorch's model.compile() function, see here and here)

I am a private Windows user, and I already had some few times where Python packages did not work as they should. Therefore I tend to use Docker containers more often 😅

[rayed-therap]

Hi @cristobalmackenzie,

After following the instructions for the dev container setup for vscode, you will be able to run jupyter notebooks directly in vscode because the jupyter extension is already installed in the container.

How I normally do this is to simply open a notebook in vscode and run a cell. VSCode will promt you to install ipykernel which is needed by vscode. Once the installation completes, you're cell should run.

If VSCode complains that is needs a python interpreter, hit Ctrl + Shift + P and start typing select interpreter. Press enter and select the /opt/conda/bin/python interpreter.

Let me know if this helps. Else we can arrange a call and maybe I can screen share.

[rayed-therap]

Here's all that I had to do to run the project in Windows using WSL2:

System Requirements:
Windows version: 11
WSL version: 2.1.4.0
Docker Desktop version: 4.28.0
Ubuntu: 22.04.4 LTS

Steps:

1. Install NVIDIA Container Toolkit as described here. NVIDIA Container Toolkit is supported as written here also because it's just an apt install and WSL2 is a full linux kernel.
2. Add runtime in Docker Desktop daemon config and restart:
   

Make sure desktop-linux builder is running:

[d-kleine]

Great, thanks - it now works with "runArgs": ["--runtime=nvidia", "--gpus=all"] 😃 I did not know that I need to add the nvidia runtime in Docker Desktop daemon config, that was new for me!

I provide the Docker engine daemon config code from above for copying, in case other readers might need that setting too:

{
  "builder": {
    "gc": {
      "defaultKeepStorage": "20GB",
      "enabled": true
    }
  },
  "experimental": false,
  "runtimes": {
    "nvidia": {
      "path": "nvidia-container-runtime",
      "runtimeArgs": []
    }
  }
}

One other thing I have noticed it that VS Code's Source Control feature does not work for me in the dev container. So, I cannot see if I have edited some files while being in the container. Do you have idea how to fix that issue?

On Windows, when Docker dev container is not running:

When dev container is running:

How can I fix that when running the dev container in VS Code that the Source Control tree using Git will be displayed too?

[rayed-therap]

Happy to hear it worked. I'll try to investigate the Git issue on WSL and get back to you.

[d-kleine]

BTW I still believe "--runtime=nvidia" and adding the nvidia runtime in Docker Desktop daemon config is not needed for Windows + WSL 2 anymore as it has been deprecated, see here. For WSL 2, the GPU support with CUDA has been merged in Docker Desktop, see here. Therefore, you don't need to specify the nvidia runtime anymore when running Docker Desktop in Windows with WSL 2. That would explain why it worked for me with "--gpus=all" only.

Please also see here the most voted answer.

[d-kleine]

@rayedbw I have found the issue why Git in the Source Control feature is not working: git will not be installed with the commands you have provided. The code in the Dockerfile should be

...
RUN apt-get update && \
    apt-get upgrade -y && \
    apt-get install -y git && \
    rm -rf /var/lib/apt/lists/*
...

1. There is no EOF command in Docker, therefore you need to use RUN and also the && \ concatenating the commands into a single line.
2. The Ubuntu package updates need to be pulled with apt-get update and then installed with apt-get upgrade -y
3. If have also removed the --no-install-recommends for the git install as this might lead to problems, please see https://askubuntu.com/questions/65081/what-are-the-implications-of-no-install-recommends-apt-get-install

See here why I went for these changes.

Please double-check if that works for you too.

[cristobalmackenzie]

Hi @rasbt, @rayedbw ! Thanks for the help ! Turns out I was only missing this:

How I normally do this is to simply open a notebook in vscode and run a cell. VSCode will promt you to install ipykernel which is needed by vscode. Once the installation completes, you're cell should run.

Everything working now. Thanks again!

[rayed-therap]

That's great to hear @cristobalmackenzie. Happy coding.