Synoros builds geometric and spectral machinery for machine learning that stays numerically exact on the hardware learning actually runs on.
The primitives are differentiable and composable, spanning Riemannian
manifolds, hyperbolic graph operations, spectral graph theory, discrete
curvature, and computational topology. They are built to hold under
float64, automatic differentiation, and GPU execution, so a model can work
on geometric structure directly instead of reconstructing it from data.
- holonomy_lib is the open PyTorch library of these primitives. GPU-native, batched-first, every result cited and checked against an established library where one exists.
- The preprint documenting it, Meaning on a Video Card: Scalable, Lossless Geometric Primitives, is at 10.5281/zenodo.20451005.