After 8 years, I rewrote my open-source PyTorch curvature library
Summary
A detailed write-up about rewriting an open-source PyTorch curvature library after eight years. The article explains the library's purpose (eigendecomposition of the Hessian and related curvature matrices), the algorithms used (Lanczos, stochastic power iteration, Hutch++, spectral density), install and usage notes, and guidance on when to use alternative operators like GGN or EmpiricalFisher, along with practical deployment considerations.