#awq

Calibration Is the Work

Round-to-nearest quantization distributes precision evenly across all weights. That is the wrong allocation. A tiny fraction of weight channels — the ones multiplied by large activations — dominate model output. AWQ finds those channels first. RTN never looks.