Suggestion
The current installation page leads with the from-source / CVMFS instructions and only mentions the Docker & Apptainer images further down. For most users — and especially on HPC clusters where Apptainer is already available — pulling the pre-built image is the fastest, most reliable way to get started.
I'd propose reorganising docs/source/installation/install.rst so that:
- The page opens with a short pointer to the Docker/Apptainer section as the recommended path.
- Docker & Apptainer Images comes first, framed as the default.
- Installing From Source follows, framed as the option for writable/editable installs or systems without a container runtime.
- The IceTray subsection under "From Source" is rewritten to point back up to the IceTray+GraphNeT image as the preferred route, with the CVMFS instructions retained as a fully-supported alternative for users who prefer working directly inside a CVMFS environment.
No technical content changes — just ordering and framing.
Motivation
- I actually didn't know about the dockers really untill this week
- I meet some people who had issues getting GraphNeT to work on their HPC clusters were using the Docker just to run stuff is probably easier
- The IceTray subsection already says "we highly recommend to instead use our existing Docker images", but a new user has to scroll past a full CVMFS recipe to find that out.
- HPC users often can't install from source on the bare cluster — for example RHEL/Rocky 8 nodes ship glibc 2.28, while
pyg_lib's libpyg.so in the PyPI wheels at data.pyg.org requires GLIBC_2.29, so an pip install -e .[torch-XX] outside of CVMFS/conda-forge tends to fail at import time. The pre-built images sidestep this entirely.
- CVMFS-based IceTray + GraphNeT setups are still fully supported and remain documented under "From Source"; this is purely a reordering/framing change.
- The pre-built images already cover both CPU and GPU and are autobuilt per release, so leading with them better matches what's actively maintained.
I have a local rewrite ready and am happy to open a PR if there's interest.
Suggestion
The current installation page leads with the from-source / CVMFS instructions and only mentions the Docker & Apptainer images further down. For most users — and especially on HPC clusters where Apptainer is already available — pulling the pre-built image is the fastest, most reliable way to get started.
I'd propose reorganising
docs/source/installation/install.rstso that:No technical content changes — just ordering and framing.
Motivation
pyg_lib'slibpyg.soin the PyPI wheels atdata.pyg.orgrequiresGLIBC_2.29, so anpip install -e .[torch-XX]outside of CVMFS/conda-forge tends to fail at import time. The pre-built images sidestep this entirely.I have a local rewrite ready and am happy to open a PR if there's interest.