Python dependency clarification

[FlyinTeller]

Hi,

you provide two different environment.yml files with very different requirements in terms of the pytorch version. From what I can gather, the _blackwell one is for newer GPU architectures and newer cuda versions.

What I am wondering: How strict are the requirements on the torch version exactly? Does the code work with >2.0?

[biochem-fan]

Blackwell GPUs (50x0) needs CUDA >= 12.9. Only new versions of PyTorch are compiled with it. This is why we specified new PyTorch for Blackwell.

On the other hand, the new PyTorch dropped support for Pascal GPUs (10x0), which are still in use at various institutions. This is why we specified older PyTorch for compatibility.

The Python codes (e.g. ModelAngelo) themselves should be compatible with both versions of PyTorch. I do not know exactly what is the minimum required version.

[FlyinTeller]

thank you for the clarification.

Can you say something about the other dependencies in https://github.com/3dem/relion/blob/master/environment_blackwell.yml ? Are the other ones as restrictive as written there:

- tqdm==4.65.0
- mrcfile==1.4.3
- loguru==0.7.0
- scikit-learn==1.3.0
- umap-learn==0.5.3
- matplotlib==3.7.2
- morphosamplers==0.0.13
- pydantic==1.10.18
- napari[all]==0.4.18
- tsnecuda==3.0.1
- PyQt5==5.15.9
- typer==0.9.0
- click<8.2.0
- biopython==1.81
- fastcluster==1.2.6
- seaborn==0.12.2
- dill==0.3.7
- numpy==1.26.1
- scipy==1.11.2

Can you give some guidance which of those could be relaxed and to which degree?

[biochem-fan]

I have no idea. I know NumPy >= 2.0 has many breaking changes. Probably others are not so restrictive. But the only way is to try yourself if you really want to know.

[huwjenkins]

@biochem-fan do you know if the RELION team has plans to update the dependencies for the RELION-5.1 release?

relion/environment.yml

Line 6 in 45f0a30

- python=3.10

Python 3.10 will be EoL 31 Oct 2026.

[biochem-fan]

@huwjenkins If someone tests a newer version of Python and packages and send us a pull request, I am happy to merge it. The packages must remain compatible with Pascal GPUs.

I myself am very skeptical about some institute policies banning software that reached EOL. The risk must be evaluated on use cases. If you run a Web server that receives files from untrustable users and process them with RELION, yes, you must be very careful. Probably RELION itself has many security flaws (in STAR parser, MRC parser, EER decoder etc), even if you harden the Python side. But if you use the conda environment for yourself alone, I wouldn't worry.