I've mostly used venv with pip or Conda for isolated Python environments, but my workflow and knowledge are several years out of date. My current Linux setup includes mise, and I'm trying to decide whether I should learn a newer tool such as uv or use something more Conda-like.
I want a setup that can replace the parts of pip and Conda I rely on, while remaining explicit and predictable. CUDA support for JAX, PyTorch, and TensorFlow is especially important. Ideally, the Python packages and compatible CUDA-related user-space libraries would be handled automatically, because manually installing NVIDIA and CUDA components has caused me a lot of trouble in the past. The current JAX CUDA extra installed through pip has worked surprisingly well.
Should I invest in uv or mise, or is it better to stick with pip because that is what most machine-learning projects officially recommend? Would Pixi be a better fit if I need the Conda ecosystem?
5 Answers
For a normal Python project, uv is probably the best default today. It is a fast standalone binary, so you can bootstrap an environment without having Python already installed, and it supports a pip-compatible interface as well as lockfiles and project management. You can still install JAX, PyTorch, or TensorFlow wheels and their documented CUDA extras through it.
I moved from Poetry and pyenv to uv and have not needed to go back. The workflow is simpler and much faster, especially for disposable environments and reproducible project installs. For CUDA packages, I would follow the installation instructions from the specific ML project rather than trying to manually assemble CUDA packages.
If you specifically need a Conda-style environment with compiled dependencies, compilers, and packages from conda-forge, use Pixi. It handles the broader environment and can use uv for Python packages, while also providing lockfiles and task running. For CUDA-heavy machine-learning work, that may be a more comfortable fit than a PyPI-only workflow.
That is the main tradeoff: uv is excellent when PyPI wheels cover everything, while Pixi is more appropriate when you need non-Python libraries or the wider Conda ecosystem.
A practical combination is mise for selecting general tool and language versions, then uv inside each Python project for dependencies and virtual environments. mise and uv solve different problems, so you do not necessarily have to choose between them.
The fact that ML documentation still says pip does not make uv incompatible with it. uv provides a pip-style command and generally installs the same wheels. Start with uv for a small project; if you run into dependencies that only exist or work reliably through conda-forge, move that project to Pixi instead.
The important distinction is that uv can install the Python-side CUDA libraries, but you still need a compatible NVIDIA driver on the system. It does not eliminate every piece of system-level GPU setup.