Source
Frantar et al. (2022), LIT-081 — ARXIV-2210.17323.
The mechanism
Round-to-nearest treats each weight as independent, which it is not: the layer’s output error depends on the interaction between weights, and a small rounding error in one column can be absorbed by adjusting the columns that have not been quantized yet. GPTQ does exactly that, ordering the work column-wise and using approximate second-order (Hessian) information from a small calibration set to decide the compensation.
What it buys: 3- and 4-bit weight quantization of models up to 175B, with accuracy close to the uncompressed baseline, in around four GPU-hours — a one-shot procedure with no retraining and no gradient steps. Before this, getting to 4 bits at that scale meant quantization-aware training.
The method and these figures are LIT-081’s, and the recommendation starts there: Frantar et al. proposed column-wise rounding with Hessian-guided error compensation as the replacement for round-to-nearest, and reported the 3- and 4-bit results up to 175B from a one-shot pass over a small calibration set.
The scope worth holding onto
This is weight-only, and the win is memory and bandwidth rather than arithmetic: it makes a model fit and makes decode faster because decode is bandwidth-bound, and it does not make a compute-bound prefill faster. It also needs a calibration set, so it is a procedure with an input rather than a pure transformation, and a calibration distribution unlike the serving distribution is a real failure mode.
Filed as the approach, not the artifact
Error-compensating one-shot PTQ with a calibration set is what converged; the specific search and ordering have been improved on repeatedly since. The record’s other quantization entries sit at different altitudes —
SOTA-160 treats the quantization plan and the pretraining budget as one decision, and SOTA-163 is about the number format rather than the rounding procedure.
The live alternative at the same altitude is SOTA-356 (AWQ), which protects activation-salient channels by scaling instead of compensating errors. Its authors ran the head-to-head against this method on LLaMA and Llama-2 at 7B to 70B and report lower perplexity than GPTQ with or without its reordering trick; and on OPT-6.7B at INT3-g128, calibrating on one Pile subset and evaluating on another cost GPTQ 2.3–4.9 perplexity against AWQ’s 0.5–0.6. That is the calibration-distribution failure mode above, measured.
Known implementations
- the default weight-only path in most open-weight serving stacks