Add torch 2.9/2.10/2.12 support via torch-geometric 2.8

[sevmag]

Summary

Adds torch-29, torch-210, and torch-212 extras (torch 2.9.1 / 2.10.0 / 2.12.1), enabled by simultaneously moving these extras to torch-geometric==2.8.* — plus fixes for two latent graphnet bugs the new stack exposed.

Why torch-geometric 2.8 is the key

torch-geometric 2.8 rewrote knn_graph/radius_graph to run on pyg-lib>=0.6.0 (torch.ops.pyg.knn/radius) and dropped the torch-cluster backend — and pyg-lib >=0.6.0 wheels exist for all torch >=2.8. This resolves the reason for the existing torch-geometric<2.8 cap, which stays in place for the older extras (their torch versions have no compatible pyg-lib). The new extras therefore drop torch_cluster (superseded by pyg-lib) and torch_spline_conv (never imported by graphnet).

Wheel availability at data.pyg.org (why the ceiling is 2.12.1)

torch	pyg_lib	torch_scatter/sparse	torch_cluster	torch_spline_conv
≤2.10	✓	✓	✓	✓
2.11	✓	✓	✓	—
2.12.x	✓	✓	—	—
2.13/2.14	✓	—	—	—

torch_scatter/torch_sparse are still imported directly by graphnet models, so 2.12.1 is the highest reachable version without migrating those imports to torch_geometric.utils.scatter/native ops — left for a follow-up.

Latent bugs exposed by the backend swap (fixed here)

The first CI run failed two icetray-job tests. Both turned out to be pre-existing graphnet bugs that the knn-backend change surfaced, reproduced and verified in the icecube/icetray:icetray-devel-v1.13.0 container:

1. cleaning_module._split_pulsemap_in_dom_types: iterating an I3Map yields OMKeys (not key/value pairs), and OMKey is itself indexable (om_key[0] is the string number), so the P[0]-style unpacking raised TypeError: Invalid index type on the first non-empty cleaned pulsemap. The loop body had never executed in CI — the QUESO test model cleaned away every pulse in the tiny test file, and pyg-lib's different knn tie-breaking now lets pulses survive the threshold. Fixed by iterating items(); correct on the old stack too.

2. KNNEdges/RadialEdges on empty events: pyg-lib's knn/radius kernels segfault on empty input when batch is None (knn_graph(torch.rand(0, 3), k=8) reproduces it; with an explicit batch tensor it is fine — worth an upstream pyg-lib report). An empty event inside a DataLoader worker is exactly this call, killing the worker (DataLoader worker ... killed by signal: Segmentation fault). Not icetray-specific — the ERDA test just only runs in the icetray job. Fixed with an explicit empty edge_index guard, matching torch-cluster's behaviour.

Changes

• setup.py: three new extras; existing extras untouched.
• src/graphnet/deployment/icecube/cleaning_module.py, src/graphnet/models/data_representation/graphs/edges/edges.py: the two fixes above.
• .github/workflows/build.yml: icetray and macOS jobs on 2.10.0, Ubuntu examples matrix on 2.12.1/2.10.0/2.9.1. pip show torch-cluster → pip show pyg-lib (the step runs under bash -e, and torch-cluster is no longer installed).
• .github/actions/install/action.yml: extended the documented torch_version/hardware option lists (cu128/cu130/cu132).
• docs/source/installation/quick-start.html: picker offers the new versions with their CUDA flavors (verified against the published wheel indexes: 2.9.1/2.10.0 → cu126/cu128/cu130, 2.12.1 → cu126/cu130/cu132); unsupported torch/CUDA pairs are now derived from the install-command map instead of a hand-maintained blacklist. The optional jammy_flows line also gets the numpy<2 constraint from Fix IceTray CI: keep numpy<2 in the jammy_flows install resolve #914.

Validation

• Examples matrix (2.9.1/2.10.0/2.12.1), macOS, docs: green on the first CI run already.
• The two icetray failures were reproduced in the v1.13.0 container on a local cluster, root-caused as above, and both tests pass there with the fixes (alone and in CI test ordering, torch 2.10.0 + tg 2.8).
• Behavioral note: the knn backend changes from torch-cluster to pyg-lib; semantics are equivalent up to tie-breaking among equidistant neighbors.

🤖 Generated with Claude Code

https://claude.ai/code/session_014H5hSfPGXaV5aabm1g45oH