GPTQ-2D: Cubic-Time Two-Sided Adaptive Rounding for Matrix Quantization

ISTA-DASLab · hf · 2026-08-04

The paper introduces the GPTQ-2D algorithm, solving the computational complexity issue of two-sided adaptive rounding in matrix quantization.

Background: Adaptive rounding methods like GPTQ round real matrices to integers under a quadratic metric. When extended to two-sided tasks (where fixed nonsingular basis matrices act on both left and right of the residual), the traditional 1D algorithm applies but takes quartic time in the matrix dimension.
Breakthrough: GPTQ-2D produces the identical rounded matrix in cubic time. It processes entries anti-diagonal by anti-diagonal; since entries on the same anti-diagonal are independent, they can be rounded in parallel, significantly boosting quantization efficiency.

More from Infra