Post-training quantization methods are designed with a Hessian-weighted layer error measured on the calibration Hessian. Error-feedback rounders (GPTQ, LDLQ and
Quantized packages on this problem are usually compared one build per method with an item-paired test. A build is one calibration draw and encoder run. Such tes
Propagation-aware calibration (GPTAQ, QEP, CoreQ) quantizes a layer so that its output matches the clean model's output, partly cancelling error from earlier qu