[Feature Request] - Relax dependency versions

[jolespin]

dependencies = [
    "vambcore == 0.1.2",
    "numpy == 1.26.4",
    "torch == 2.6.0",
    "pycoverm == 0.6.2",
    "networkx == 3.4.2",
    "scikit-learn == 1.6.1",
    "dadaptation == 3.2",
    "loguru == 0.7.3",
    "pyhmmer == 0.10.15",
    "pyrodigal == 3.6.3",
]

I'd like to use VAMB in some of my pipelines but the current dependency specifications are not flexible. Would it be possible to provide some ranges here?

[jakobnissen]

Yeah possibly, I'll think about it.

Earlier versions of Vamb did have more lax compat versions, but I got burned by it too many times. The Python ecosystem is quite bad with not respecting semver, so there was recurrent issues with things randomly breaking whenever the dependencies got upgraded.

The issue is that it's hard to test with all possible different versions. What do other people why maintain Python packages do? Perhaps provide a range and then run CI for the lowest possible versions and the highest possible versions of all (transitive) dependencies?

[jolespin]

Yea what I typically do is allow all versions that have a particular usage. For example, if this is dependent on numpy 1.x I would use numpy >=1.0,<2.0 unless there's some particular feature that was added. With pyrodigal, a particular version introduced -j/--jobs and since I use that feature, I add that as a minimum version with no upper bound.

[jolespin]

For what it's worth, I've created a GPU enabled conda environment that uses VAMB, CheckM2, and some other Python packages in Python 3.12

https://github.com/jolespin/veba/blob/main/install/environments/VEBA-binning-prokaryotic-gpu_env.yml

There's a CPU env in that directory of the repo as well.

I've successfully integrated VAMB in VEBA for 2.5.x.