# GPTQ Algoritması: Optimal Brain Quantization + Hessian Update — RTX 4090'da 12 Dakikada Llama 8B

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-gptq-algorithm-optimal-brain-quantization
> Updated: 2026-08-21T03:28:59.130Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part X — Quantization Engineering
**TLDR:** GPTQ (Frantar et al. 2022) — LLM weight quantization standardı. Optimal Brain Quantization theory (LeCun 1990), Hessian inverse update, error compensation, group quantization. RTX 4090 + auto-gptq ile Llama 3.1 8B'yi 12 dakikada int4'e quantize et. WikiText-2 perplexity delta < %2.