BaKron Quantization Using Kronecker-Factored Hessians

BaKron is a quantization solver that uses two-sided Kronecker-factored Hessian approximations to capture correlations across output coordinates. It reduces computational complexity from O(m^2n^2) to O(mn(m+n)) for an m x n weight matrix, matching GPTQ scaling while utilizing richer curvature information.