BaKron: Efficient Quantization with Kronecker-Factored Hessians

The introduction of BaKron, a novel algorithm, aims to enhance the speed of neural network quantization by utilizing Kronecker-factored Hessian approximations. This approach improves upon earlier techniques like GPTQ, BoA, and YAQA, presenting a more efficient solver that lowers computational complexity from O(m^2n^2) to O(mn(m+n)) for a weight matrix of size m×n. This efficiency is achieved through anti-diagonal parallelism and a recursive divide-and-conquer strategy, which requires only O(m+n) sequential steps. BaKron maintains the cubic scaling of GPTQ while taking advantage of more detailed curvature information, ensuring modularity with the underlying quantization method. The research can be found on arXiv with the identifier 2608.06291.