Devcontainer Template

This is an up-to-date Devcontainer template for developing Python projects (with Git and Poetry).
It has optional support for TensorFlow and PyTorch on GPU enabled machines.

Instructions

1. Create a new repository using this GitHub template.
2. Clone the repository and open it using VSCode.
3. Run python .devcontainer/change_gpu_config.py to select your GPU mode.
4. Press Ctrl+Shift+P to open the command palette.
5. Search for Dev Containers: Rebuild and Reopen in Container.
6. Follow the instructions in the terminal to install TensorFlow or PyTorch.
7. You are now ready to start developing! 😀

Devcontainer info

Base OS

Name	Version	Image
Ubuntu	24.04	mcr.microsoft.com/devcontainers/base:ubuntu-24.04

Features

Name	Version	Image	Optional
NVIDIA CUDA	CUDA 12.6 & cuDNN 9	ghcr.io/devcontainers/features/nvidia-cuda:1	Yes
Python	3.12	ghcr.io/devcontainers/features/python:1	No
Git	Latest	ghcr.io/devcontainers/features/git:1	No
GitHub CLI	Latest	ghcr.io/devcontainers/features/github-cli:1	No

Libraries

Name	Version	Description	Optional
TensorFlow	2.17.0	AI development framework	Yes
PyTorch	2.5.0	AI development framework	Yes
Poetry	Latest	Venv-like solution with great dependency management	No

Extensions

Name	Description	ID
Python	Python language support	ms-python.python
Pylint	Static code analyser	ms-python.vscode-pylint
Black	Code formatter	ms-python.black-formatter
Jupyter	Jupyter extension pack	ms-toolsai.jupyter
Prettier	Code formatter	esbenp.prettier-vscode

Possible errors

❌ Error	✅ Solution
Shell scripts fail to run or complain about \r characters	Check if the End of Line formatting of the scripts is set to LF in VSCode.

Versioning

CUDA Toolkit	Linux Driver Version	Windows Driver Version
CUDA 12.x	>= 525.60.13	>= 527.41
CUDA 11.x	>= 450.80.02	>= 452.39

• The CUDA Toolkits have minor version compatibility with the drivers.
  This means that CUDA Toolkit 12.6 should work with a driver that's designed for CUDA 12.0 because they have the same major version.
• The CUDA Toolkits are backwards compatible with the drivers.
  This means that CUDA Toolkit 11.8 will still work with a newer driver that's designed for CUDA 12.x versions.

Latest version	Python version	CUDA
TensorFlow 2.17.0	3.9-3.12	12.3
PyTorch 2.5.0	3.9-3.12	12.4

The most recent common Python version is 3.12.

GPU accelerated containers

The hosts NVIDIA driver gets passed to the container using the NVIDIA Container Toolkit. You can validate your Container Toolkit installation by checking the Docker daemon configuration file on your server: /etc/docker/daemon.json. To spin up a GPU-accelerated container, append the --gpus=all and --runtime=nvidia arguments to your docker run command. Luckily, these arguments are already preconfigured in devcontainer.json.

The NVIDIA driver on the A5000 server has version 12.0.
A GPU-enabled container requires the NVIDIA CUDA Toolkit (contains cuFFT, cuBLAS, etc.) and cuDNN in the container itself.