Installation et utilisation d’environnements logiciels pour l’IA scientifique

Groupe de travail AI au LEGI, lundi 30 mars

Pierre Augier

Different kinds of AI

• Machine learning
• Deep learning, neural networks
• Generative IA. Transformers and co.
• …

Many different applications for science…

AI usage : remote or local

Chat LLM websites

https://chat.mistral.ai/, https://claude.ai, …

Vibe and agents

Mistral Vibe, Toad, Claude Code, Qwen Code (Alibaba), Gemini CLI, …

Open-weight models

Albert API (DINUM), HuggingFace, Ollama, LM Studio, ILAAS

Machine learning and deep learning (local and clusters)

Some applications require manual software installations

About ecology and AI financial bubble

Current AI dynamics is not sustainable: ecology, ressources, energy and finance

However,

• AI dynamics is not only slop, communication and lies

  In particular, for research activities

• Running large models locally is not the solution

• \(\neq\) AI usages, \(\neq\) impacts

  • other activities (coding, computing, experiments, travels, …)?
  • which benefit?

Software installation: good practices & good tools

• Fast moving field: recent soft and versions, apt install not adapted

• We want a bit of reproducibility: versioned lock files

• We want flexibility and ease of use

Recent tools and practices offer that: project managers

• Pixi (conda ecosystem + PyPI)
• UV (PyPI)

Warning

Avoid older, less advanced tools (conda, pip, …)

A good starting point

Note

• Python is still highly dominant in the field of AI
• With C++ in the background…
• Mojo uses Pixi

Professional Python Environment Setup: Reference Guide

(by the Python CNRS Working Group and the py-edu-fr project)

https://python-cnrs.netlify.app/edu/init/preliminaries/

Note

Pixi and UV are must-have

Examples with Pixi

Simple environnement with scikit-learn and Python 3.14

pixi init
pixi add scikit-learn
pixi shell

Pytorch

https://pixi.prefix.dev/latest/python/pytorch/