How do I use pixi and uv together

[sidmadala]

I am wondering how I can best use pixi and uv together. Previously, I have been using conda/mamba alongside pip which enables me to install pypi dependencies inside the appropriate conda environment. Should I add an instance of uv to each pixi environment I wish to create, or should I do pixi global install uv and activate pixi shell inside the environment I want to use uv in? Any comprehensive documentation for how this interop works relative to conda and pip would be much appreciated. The reason I ask this is because I need to use uv for #3730 via --no-build-isolation alongside options when installing specific packages.

NOTE:

I have tried the shell method already with a global uv installation and it seems to create a .venv folder in my project even though these packages should just be installed in the .pixi environment located in the same directory. It appears that I can delete the .venv and still use the pypi libraries I installed via uv. Is this recommended behavior?

[lucharo]

I wonder the same!!

[maurosilber]

It is discussed in the docs: https://pixi.sh/latest/python/tutorial/#managing-both-conda-and-pypi-dependencies-in-pixi

• pixi add <package> adds a conda package (in the dependencies table),
• pixi add --pypi <package> adds a PyPI package (using uv as a library, in the pypi-dependencies table).

Any package installed without pixi (for example, using uv directly as a binary) is not recorded in the pixi.lock file and pixi would not know it is there. After installing the environment, you could do [uv] pip install <other-package>, but it is not recommended.

[lucharo]

That's great, thank you so much @maurosilber, that's more clear now!

Regarding then publishing to pypi, even projects that might include conda dependencies, is pixi-pack the default solution? I work in an environment full of bioconda dependencies but I still rely on publishing my projects as pypi packages for sharing with other colleagues and CI workflows.

https://pixi.sh/latest/deployment/pixi_pack/#mirror-and-s3-middleware

https://pixi.sh/latest/build/python/#lets-get-started

[maurosilber]

pixi-pack is to pack an environment, not create a (conda) package. It is more similar to creating a Docker container.

pixi build is to build a conda package, which can then be uploaded to https://prefix.dev or https://anaconda.org. If you want to publish to conda-forge or bioconda, they do not support pixi build yet, as far as I know.

If you want to publish to PyPI, you need to create a Python package, which is shown before the Adding a pixi.toml section. To build and publish it to PyPI, I use uv in a GitHub workflow (which I install in an environment using pixi)

[lucharo]

In the case of wanting to publish to pypi would we then need to maintain both a pixi.toml and a pyproject.toml? How would these two interact? For instance if I have some [pypi-dependencies] in pixi.toml and some dependencies in pyproject.toml?

[maurosilber]

You can merge pixi.toml into pyproject.toml by adding a tool.pixi prefix to every table: https://pixi.sh/latest/python/tutorial/#pixitoml-and-pyprojecttoml

pyproject.toml:

[project]
dependencies = []  # PyPI dependencies (for both other tools and pixi)

[tool.pixi.dependencies]
# conda dependencies

[tool.pixi.pypi-dependencies]
# also PyPI dependencies but only read by pixi

So, you don't need to maintain two sources of pypi-dependencies, just put them in project.dependencies (if they are runtime dependencies for your package).

[lucascolley]

See https://github.com/data-apis/array-api-extra/blob/main/pyproject.toml and https://github.com/data-apis/array-api-extra/blob/main/.github/workflows/cd.yml for an example of this with pyproject.toml and some actions for deploying to PyPI.