Restricted dependencies

[kyriakosschwarz]

Hi, this new version of RXNMapper looks great!

However when trying to run on the newest GPUs some issues arise due to dependency restrictions (and especially numpy for example).

Would it be possible to make those dependency restrictions much more flexible? It would be greatly appreciated!

[yvsgrndjn]

Hi @kyriakosschwarz thank you for the feedback!

Does the problem arise with Python 3.11? I tested it on a nvidia 50 series and worked fine, but I might be missing something.

Once it is clearer where the actual conflict is happening, I'll see if I can relax the constraints. If it requires modifying the original RXNMapper class (which is unmodified in this version, we added things around it), I'll see how we can proceed! From what I recall the numpy restriction comes from there.

Let me know!

[kyriakosschwarz]

Hi @yvsgrndjn thanks for the quick reply! I am using python 3.12+, that's why I run into issues. I cannot use older versions due to other packages. Here's also an LLM analysis:

The main blocker is pyproject.toml: it pins numpy>=1.20,<1.24. With binary wheels only, the repo resolves on Python 3.11, but fails on Python 3.12, 3.13, 3.14, and 3.15 because NumPy 1.23.x has no usable wheels there. If source builds are allowed, uv can still “resolve” Python 3.12+ to numpy==1.23.5, but that is exactly the brittle path: old NumPy source builds against new Python.

There are also stale ML pins/metadata:

transformers==4.46.3 in pyproject.toml is old relative to current 5.12.1. The bundled ALBERT configs still load under both 4.46.3 and 5.12.1 without warnings, but full runtime compatibility with Transformers 5 would need test coverage.
setup.cfg is inconsistent/stale: it has torch>=1.5.0,<2.5, transformers>=4.0.0, old Python classifiers, and package data that only lists the original model. Modern builds likely use pyproject.toml, but the stale file is a packaging hazard.
CUDA itself is probably okay for the repo’s own code path in core.py, since it just does torch.cuda.is_available() and moves the model to CUDA. With uv, current Torch resolves on Python 3.14 for cpu, cu126, and cu130 as torch==2.12.1; cu128 resolves to torch==2.11.0.

So the practical compatibility read is:

Best current safe target for this repo as-is: Python 3.11.
Python 3.12+ as-is: not clean because of numpy<1.24.
Python 3.14 + modern NumPy/SciPy/Pandas/Torch/RDKit: ecosystem supports it, but this repo’s constraints do not.

Thanks again for your support!

[yvsgrndjn]

Yeah if you have to use 3.12+ I recommend trying to remove the constraint from the pyproject.toml and see if something breaks. In the original RXNMapper they did not enforce a numpy version, but I think you get errors later if you have numpy >1.23
I am planning on looking into this later, but for now please try that and keep me updated!
PRs are also welcome in case you decide to update it yourself :)