IST-DASLab / gptq

Code for the ICLR 2023 paper "GPTQ: Accurate Post-training Quantization of Generative Pretrained Transformers".
https://arxiv.org/abs/2210.17323
Apache License 2.0
1.81k stars 145 forks source link

Use modified Cholesky decomposition instead of regularized Cholesky #41

Open jiahao opened 11 months ago

jiahao commented 11 months ago

Hello, I just wanted to point out that the problem you are trying to avoid with regularization of indefinite Hessians has been encountered in the numerical analysis and optimization literature, and the class of algorithms goes by the name modified Cholesky, with cost $O(n^2)$ which is asymptotically irrelevant.