How do folks build wheels for different groups of dependencies using a single pyproject.toml?

[rbavery]

Until dependency groups are supported, what tools handle building wheels from multiple overlapping dependencies well?

Right now I'm resorting to commenting/uncommenting dependencies from my poetry main group to build a cpu or gpu version of my package which refer to different sources.

[tool.poetry.dependencies]
python = "^3.10"
geopandas = "^0.11.1"
memory-profiler = "^0.61.0"
....

[tool.poetry.dev-dependencies]
....

[tool.poetry.group.dev.dependencies]
hypothesis = "^6.97.4"
poetry-plugin-export = "^1.6.0"

# TODO poetry can't install group dependencies without resolving all groups???
# poetry also can't build wheels using dependency groups
[tool.poetry.group.gpu.dependencies]
torch = {version = "^2.3.0.dev20240311+cu121", allow-prereleases = true, source = "pytorch-gpu-nightly"}
torchvision = {version = "^0.18.0.dev20240311+cu121", allow-prereleases = true, source = "pytorch-gpu-nightly"}
pytorch-triton = {version = "^3.0.0+a9bc1a3647", allow-prereleases = true, source = "pytorch-gpu-nightly"}

# [tool.poetry.group.cpu.dependencies]
# torch = {version = "^2.2.0", allow-prereleases = true, source = "pytorch-cpu-stable"}
# torchvision = {version = "^0.17.0", allow-prereleases = true, source = "pytorch-cpu-stable"}

[[tool.poetry.source]]
name = "pytorch-gpu-nightly"
url = "https://download.pytorch.org/whl/nightly/cu121"
priority = "explicit"

[[tool.poetry.source]]
name = "pytorch-cpu-stable"
url = "https://download.pytorch.org/whl/cpu"
priority = "explicit"

I'd like to avoid unecessarily duplicating my project.toml and having to make the same edits in two different places as my project evolves.

Instead I'd like to do something like

rye build --with-groups gpu or rye build --with-groups cpu and build the wheel so that I only resolve the main dependencies + whatever group(s) I specify