GraphEm RAPIDS

GraphEm RAPIDS provides one implementation of the corrected GraphEm algorithm: a device-parameterized Torch normalized-Laplacian start followed by CuPy/cuVS force refinement on CUDA.

Production runs use device="cuda" and fail if Torch CUDA is unavailable. The device parameter applies only to the shared Torch spectral function; GraphEmbedder always requires the CUDA/RAPIDS stack for graph storage, midpoint search, forces, and updates. Diagnostic CPU spectral selection warns and does not constitute a CPU layout backend.

The package validates graph invariants, spectral residuals, orthogonality, midpoint identities, and finite layout state. A failed check raises an error rather than selecting another solver or algorithm.

User guide

• Quick start
  • Requirements
  • Device policy
  • Adjacency input
  • Edge-list input
  • Failure and diagnostics
• API reference
  • GraphEmbedder
  • Graph generators
• Graph Generators
  • Random Graphs
  • Scale-Free Graphs
  • Small-World Graphs
  • Community Structures
  • Bipartite Graphs
  • Geometric Graphs
  • Tree Structures
  • Complete Example