BaKron: Efficient Quantization with Kronecker-Factored Hessians
BaKron makes two-sided Hessian-aware quantization scale like GPTQ instead of blowing up in the vectorized weight space.
The paper targets adaptive rounding methods for neural network quantization that use Kronecker-factored Hessian geometry. Its solver combines anti-diagonal parallelism with a recursive divide-and-conquer construction. For an `m x n` weight matrix, the authors say it cuts work from `O(m^2n^2)` to `O(mn(m+n))` while using `O(m+n)` sequential steps. They also describe BaKron as modular across base quantizers and Hessian estimators, with benchmarks and experiments included. ArXiv · AI/CL/LG's note
The paper targets adaptive rounding methods for neural network quantization that use Kronecker-factored Hessian geometry. Its solver combines anti-diagonal parallelism with a recursive divide-and-conquer construction. For an `m x n` weight matrix, the authors say it cuts work from `O(m^2n^2)` to `O(mn(m+n))` while using `O(m+n)` sequential steps. They also describe BaKron as modular across base quantizers and Hessian estimators, with benchmarks and experiments included. ArXiv · AI/CL/LG's note
score 4