pixi shell should only pull the requested environment, not all environments.

[sebastian-sz]

Problem description

TL;DR pixi shell pulls ALL environments, even when only the default is requested.

I'm rolling a Pytorch CPU / GPU project. linux-64 uses GPU and CPU, everything else is CPU. Please find the pyproject.toml below:

[project]
authors = [{name = "Sebastian Sz", email = "***@***.com"}]
dependencies = ["jaxtyping>=0.2.34,<0.3"]
description = "Add a short description here"
name = "torch-proj"
requires-python = ">= 3.9, <3.13"
version = "0.1.0"

[build-system]
build-backend = "hatchling.build"
requires = ["hatchling"]

[tool.pixi.project]
channels = ["conda-forge", "pytorch"]
platforms = ["linux-64", "osx-arm64", "win-64"]

[tool.pixi.dependencies]
pytorch = {version = "2.4.1", channel = "pytorch"}
torchvision = {version = "0.19.1", channel = "pytorch" }

[tool.pixi.pypi-dependencies]
torch-proj = { path = ".", editable = true }

[tool.pixi.tasks]

[tool.pixi.environments]
cuda = ["cuda"]

# CUDA setup
[tool.pixi.feature.cuda]
platforms = ["linux-64"]
channels = ["nvidia/label/cuda-12.4.0", "nvidia", "conda-forge", "pytorch"]

[tool.pixi.feature.cuda.dependencies]
pytorch-cuda = {version = "12.4.*", channel = "pytorch"}

The issue is: when running pixi shell on Linux, the cuda environment gets downloaded and installed, even though I did not request cuda. This is kinda bloaty, e.g. when I want to run CI on Linux on the CPU there is no reason to pull the Nvidia's packages (which are quite heavy).

Ideally, we would pull cuda only when requested, e.g:
pixi shell <- downloads CPU only packages
pixi shell -e cuda <- downloads CUDA only packages

Rather then have everything possible downloaded, regardless of the environment.

Pixi version: 0.30.0

[sebastian-sz]

The vice versa also holds: I shouldn't have to pull cpu environment if I plan to use cuda only.

If there is any change I can do in my workflow, please don't hesitate to let me know!

[ruben-arts]

The problem your having is caused by our requirement to install the conda environments to solve for the pypi dependencies. What you require for this to work is the feature requested in: #1046

What you could try for now is a pip install --editable . task, instead of the following configuration:

[tool.pixi.pypi-dependencies]
torch-proj = { path = ".", editable = true }

This would avoid the installation to only when you need it instead of requiring the environment at solve time.