Update conda env & remove default channel

[Zethson]

Dear team,

could we please update the environment https://github.com/NVIDIA-BioNeMo/KERMT/blob/main/environment.yml to use more latest versions of dependencies?
Additionally, we have to drop the default channel as it requires an Anaconda license that not all customers have.

Thanks!

[sveccham]

Hi @Zethson,
We have updated remove the Anaconda defaults channel and use a miniforge installed instead. All the required packages are available on conda-forge.
Are there certain dependencies that are particularly troublesome? We would need to validation model performance again before updating RDKit version.

[Zethson]

Thanks! So the way this is commonly done is that there's an associated reproducibility repository or just yaml file that hosts pinned versions that were used for the paper. However, the user facing versions are unpinned. The user facing yaml should looking something like

dependencies:
  - python=3.14.7
  - pytorch-gpu
  - rdkit
  - cuik_molmaker
  - descriptastorus
  - optuna
  - scikit-learn
  - huggingface_hub
  - protobuf

All developer dependencies are not user facing and must be removed (like pytest, ...). Not sure whether protobuf and huggingface_hub are user facing

[Zethson]

I don't think this has been correctly solved.

1. There's still pins here: https://github.com/NVIDIA-BioNeMo/KERMT/blob/main/environment.yml
2. Python 3.11 is pretty much EOL. Spec-0 will recommend 3.13+ in October.
3. There's still developer facing dependencies in the yaml.

If you add me to the repository I can make a PR.