pypi + conda deps monorepo best practices

[pablovela5620]

Problem description

I have a lot of different code spread across a bunch of repos, and I'm starting to think that maybe a monorepo might be a better setup for me. There are a few discussions about how to use Pixi in this way.

#387
#3636

and also something that it seems that uv supports in a cargo like manner

https://docs.astral.sh/uv/concepts/projects/workspaces/

As I understand, this will basically work for PyPI dependencies only in the dependencies = []
and not those found in [tool.pixi.pypi-dependencies] or [tool.pixi.dependencies]

This is sort of what I'm thinking, such that the dev/gpu-related are the same for all subpackages.

pvlabs-mono/
├─ README.md
├─ pixi.toml              # Pixi workspace manifest (single source of truth)
├─ pixi.lock              # Single lockfile for entire workspace
├─ packages/
│  ├─ simplecv/
│  │  ├─ pyproject.toml   # package metadata + CLI entrypoints
│  │  ├─ simplecv/        # source code
│  │  ├─ tools/           # package-specific scripts
│  │  ├─ tests/
│  │  └─ data/
│  ├─ rtmlib/
│  │  ├─ pyproject.toml
│  │  ├─ rtmlib/
│  │  └─ tests/
│  ├─ monoprior/
│  │  ├─ pyproject.toml
│  │  ├─ monoprior/
│  │  └─ tests/
│  └─ wilor-nano/
│     ├─ pyproject.toml
│     ├─ wilor_nano/
│     └─ tests/
└─ .github/workflows/ci.yml

[workspace]
name = "pvlabs-mono"
channels = ["https://prefix.dev/conda-forge"]
platforms = ["linux-64", "osx-arm64", "win-64"]

[dependencies]
python = "==3.12"

# ---- Install all member packages via path ----
[pypi-dependencies]
simplecv   = { path = "packages/simplecv",   editable = true }
rtmlib     = { path = "packages/rtmlib",     editable = true }
monoprior  = { path = "packages/monoprior",  editable = true }
wilor-nano = { path = "packages/wilor-nano", editable = true }

# ---- GPU Feature ----
[feature.gpu.system-requirements]
cuda = "12.8"

[feature.gpu.dependencies]
pytorch-gpu  = "*"
cuda-version = "12.8.*"
torchvision  = "*"
xformers     = "*"

# ---- Dev Feature ----
[feature.dev.pypi-dependencies]
pytest = "*"
ruff   = "*"
mypy   = "*"

# ---- Environments ----
[environments]
gpu = { features = ["gpu", "dev"], solve-group = "base" }
cpu = { features = ["dev"],        solve-group = "base" }

My questions are

1. any plans to make this work with conda deps
2. can I make tasks work in each pyproject.toml, as I understand if I have [tool.pixi.task] in the pyproject then the top level pixi.toml wouldn't get picked up
3. Are there best practices here? I don't know if this top-level toml is the right/recommended approach

[ruben-arts]

Hi @pablovela5620, this is very much inscope of what we're building.

1. any plans to make this work with conda deps

We currently have the ability to have packages in subdirectories and you can add them to the main workspace, follow the documentation for more info

2. can I make tasks work in each pyproject.toml, as I understand if I have [tool.pixi.task] in the pyproject then the top level pixi.toml wouldn't get picked up

We're planning to add this to the package table, it would be part of this proposal. (last paragraph): #4721

3. Are there best practices here? I don't know if this top-level toml is the right/recommended approach

Feel free to start trying the pixi-build workflow based on the documentation but expect big changes soon. When it's out, we'll pitch it through social media and discord.