n8programs

Full implementation w/ optimization and custom vjp for smol kernels, all, again, yoinked from pytorch: https://gist.github.com/N8python/b3e24a4f88efa52bdd81a8762b7a7238. For two 1024x1024 matrices initialized with randn(0, 0.1), the provided matrix_exp implementation diverges from...

For two 1024x1024 matrices initialized with randn(0, 1) - intentionally made to have diverging eigenvalues - the average difference, as a percent of the maxmimum element in the torch computation,...

Fundamental algorithm from: https://raw.githubusercontent.com/pytorch/pytorch/7f18ef14c1fed4e4376a75d626d98ba3c074809c/aten/src/ATen/native/LinearAlgebra.cpp

Agreed. But how would I go about doing that - given that the if on the norm is neccessary to choose the correct polynomial, and the scale factor - which...