pixi & vllm for layered installations

[geoHeil]

How can I get vllm to build on osx for MLX?

git clone https://github.com/vllm-project/vllm.git
cd vllm
pip install -r requirements-cpu.txt
pip install -e .  

After setting up the normal python environment these requirements need to be installed as a separate step. Would you encapsulate these in a setup task? Or is there a better approach? I.e. similar to the normal git based pypi packages - but where I can specify the path to the specific pyproject toml/requirements.txt

[geoHeil]

Answer from @ruben-arts on discord Argh you ran into a pickle. We currently don't have support yet of multi layer install/solves in pypi.

You can just run the pip commands in pixi env. But because of an inconsistency between the vllm dependencies and the pytorch dependencies on conda-forge there is currently no easy workaround.

I was trying the following. But keep running in snags.

[workspace]
channels = ["https://prefix.dev/conda-forge"]
platforms = ["osx-arm64"]

[tasks]

[pypi-options]
no-build-isolation = ["vllm"]

[dependencies]
numba = ">=0.61.2"
packaging = ">=24.2,<26"
setuptools = ">77"
setuptools-scm = "*"
pytorch = "==2.7.1"
torchaudio = ">=2.7.1"
torchvision = "==0.22.0"
datasets = ">=3.6.0,<4"
cmake = ">=3.26.1"
ninja = "*"
wheel = "*"
jinja2 = ">=3.1.6"
regex = "*"

[pypi-dependencies]
vllm = { path = "./vllm" }