7 / 9 Quantization & Low Precision

Second-Order Cleanup

GPTQ quantizes a weight matrix column by column and, after rounding each column, compensates for the introduced error by:

A

Rounding stochastically so the expected value of each weight is preserved

B

Only using calibration activation statistics, with no Hessian computation involved

C

Applying a straight-through estimator and re-running backprop over the whole model

D

Using second-order (Hessian) information from calibration data to update the remaining unquantized weights so they offset the error just introduced

E

Recomputing a single global scale factor once the entire layer is quantized