Robust Ultra Low-Bit Post-Training Quantization via Stable Diagonal Curvature Estimate
DASH-Q is proposed, a robust PTQ framework using diagonal Hessian approximation and iterative weighted least squares, which outperform other PTQ baselines in ultra low-bit regime and improves zero-shot accuracy by 7.01% on average and up to 14.01% over the strongest baselines.
Jaemin Kim, Sungkyun Kim, Junyeol Lee et al.
· EuroMLSys@EuroSys · 0 citations