Numpy <2 requirements conflicts with mattersim

[awvwgk]

Describe the bug
Mattergen requires numpy<2 while depending on mattersim, which changed its constraint to numpy>=2 recently.

Expected behavior
Have the possibility to install latest mattergen and mattersim together in the same environment.

[hfzxlzy]

Hi, I saw your question and it happens to be related to something I have been working on recently.

I have been trying to deploy the latest MatterSim together with MatterGen in the same environment, so I wanted to share my setup in case it is helpful.

My environment looks like this:

Python: 3.10
numpy: 1.26.4
mattergen: 1.0.3 (editable install, with minor dependency adjustments in pyproject.toml)
mattersim: 1.2.0 (installed via pip install --no-deps mattersim, so only the core package was installed without pulling in its declared dependencies)
torch: 2.8.0+cu128

In this setup:

• MatterGen runs normally
• MatterSim 1.2.0 also runs normally under numpy<2
• I tested only the inference-related features (prediction / relaxation / phonon), and all of them worked correctly
• I have not tested the training pipeline, so I cannot confirm whether training works with numpy<2

Since my MatterGen installation includes several small local modifications, I am listing the relevant changes here:

• I commented out all content under [tool.uv.sources] and [[tool.uv.index]] (the upstream torch/cu118 and PYG wheel source definitions)
• I modified the Linux torch dependencies from:
  torch==2.2.1+cu118 torchvision==0.17.1+cu118 torchaudio==2.2.1+cu118
  to match my local environment:
  torch>=2.2.1 torchvision>=0.17.1 torchaudio>=2.2.1

These adjustments were made only to match my local CUDA/PyTorch setup and do not reflect upstream defaults.

Still, I hope this information can help clarify the compatibility range.