Skip to content

uv + PyTorch + PyG on OSC — Version Compatibility

This stack has a three-way version coupling that must stay in sync. Getting any axis wrong causes segfaults (not import errors — silent C++ ABI mismatches).

The Constraint Triangle

PyTorch version ←→ PyG extension wheels (torch-scatter, torch-sparse, torch-cluster)
      ↕
  CUDA version
  1. PyTorch from PyPI bundles NVIDIA libs automatically. Do NOT use the download.pytorch.org/whl/cu* index.
  2. PyG extensions compiled against specific torch+cu combo. Wheels at https://data.pyg.org/whl/torch-{VERSION}+cu{CUDA}.html. Only C++ extensions need the flat index.
  3. PyTorch version on PyPI ≠ PyG's torch version tag. PyPI may ship torch 2.10.0 but PyG only has wheels up to 2.8.0. Mismatched versions segfault at runtime.

Current Pinned Versions (2026-03-04)

Component Version Source
Python 3.12.4 OSC module load python/3.12
PyTorch 2.8.0 (bundled cu128) PyPI (default index)
torch-scatter 2.1.2+pt28cu128 data.pyg.org/whl/torch-2.8.0+cu128.html
torch-sparse 0.6.18+pt28cu128 same flat index
torch-cluster 1.6.3+pt28cu128 same flat index
pyg-lib 0.6.0+pt28cu128 same flat index
torch-geometric 2.7.0 PyPI
RAPIDS Not integrated — removed. Single uv env.

Traps to Avoid

  • torch>=2.6 without upper bound: resolves to latest with no PyG wheels. Always pin <next_major.
  • requires-python = ">=3.10" without upper bound: uv resolves for 3.13+ where PyG wheels may not exist. Use <3.13.
  • uv-managed Python downloads: Standalone Python builds from uv can segfault on OSC's RHEL 9. Always use OSC's system Python via uv venv --python /apps/python/3.12/bin/python3.
  • [tool.uv] find-links for PyG: Use [[tool.uv.index]] with format = "flat" and explicit = true instead.
  • OSC CUDA modules (cuda/12.6, cudnn/8.x): Not needed when torch comes from PyPI. Only load for RAPIDS or custom CUDA code.