Zenaique

Match each post-training quantization method to what it protects against

Match pairs·Hard·4.0 · 0·~2 min·Asked atHugging FaceNVIDIA
Attempt it

Drag each answer to line up with its matching prompt

AWQ

No calibration in the simple case: relies on FP8's wider dynamic range vs INT8

GPTQ

You need INT4 weight only quant for memory bound decode on a 70B class model

SmoothQuant

Migrates activation outliers into the weight side so W8A8 INT8 quant stays accurate

FP8 PTQ (cast and go)

Hessian aware rounding that minimizes second order reconstruction error layer by layer

Pick AWQ or GPTQ when…

Protects salient (high magnitude) weight channels via per channel scaling before INT4 rounding

TL;DR

Each quantization method fixes a different failure mode: GPTQ minimizes second-order error, AWQ protects salient weight channels, SmoothQuant migrates activation outliers, and FP8 leans on wider dynamic range.

Memory aid
Sign in to see the mnemonic that makes this stick.
Easy to grasp

Imagine compressing a detailed photo into far fewer colors. If you pick colors blindly, the important bright spots turn into ugly blotches. Different methods fix this differently. One studies which pixels matter most and rounds those carefully (GPTQ). One notices a few super-bright pixels dominate and scales them up before compressing so they survive (AWQ). One realizes the brightness problem actually lives in the lighting, not the photo, and shifts the glare onto the photo where it is easier to handle (SmoothQuant). And one just uses a color format that already spans a wider brightness range, so no special trick is needed (FP8). Same goal, lower bits, but each guards against a different way the picture would otherwise get wrecked.

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Post-training quantization is the production lever that lets a 70B model fit and serve on hardware that could never hold it in bf16. The decode phase of LLM inference is memory-bandwidth-bound: every generated token streams the weights and the KV cache out of HBM. Halving the bytes per weight roughly doubles the achievable decode throughput, which is why low-bit quantization is not a niche trick but a default in 2026 serving stacks.

The difficulty is that you cannot just round. Transformer weights and activations have heavy-tailed distributions with rare but enormous outliers, and uniform rounding to four or eight bits destroys exactly the values that carry the most signal. Each method in this matching question was invented to neutralize a specific way that naive rounding fails. Getting the pairs right means understanding two orthogonal questions for each method: what tensors does it quantize, and which distortion does it fight.

This deep dive walks through all four, places them on the weights-only versus weights-and-activations axis, explains the failure mode each one targets, and finishes with the decision rule the question's fifth pair is testing.

Why low-bit quantization is hard: outliers and salient channels

The naive picture of quantization is a uniform grid: take a tensor, find its min and max, and map the range onto a fixed set of integer levels. With four bits you have sixteen levels, with eight bits you have 256. This works when values are roughly uniform. Transformer tensors are not. A handful of weight channels and activation dimensions carry magnitudes orders larger than the rest.

Those outliers force a brutal tradeoff. If you set the quantization range wide enough to represent them, the step size becomes so coarse that ordinary values collapse to a few levels and lose all detail. If you clip the outliers to keep a fine step size, you discard the very dimensions that dominate the output. Either way, plain rounding wrecks accuracy. The error is not random noise; it is concentrated in exactly the values that matter most.

There are two distinct outlier problems, and that distinction is the whole game in this question. On the weight side, certain channels are salient: rounding them coarsely hurts the output disproportionately. On the activation side, certain feature dimensions blow up unpredictably at runtime, and they are also data-dependent, so you cannot bake a fix into the weights without thought. Different methods attack different sides, and that is why they pair with different descriptions.

There is also a bit-width axis layered on top. INT8 is forgiving enough that weights alone rarely need fancy tricks, so the INT8 story is mostly about activations. INT4 halves the levels again, so even weights now need protection. FP8 changes the rules entirely by trading some mantissa precision for exponent range. Keeping these three regimes straight is what separates a confident answer from a guess.

GPTQ: second-order, layer-by-layer reconstruction
AWQ: protect the salient weight channels
SmoothQuant: migrate activation outliers onto weights
FP8 cast-and-go and the INT4 decision rule
Sign in to unlock the full deep dive.

Situations where this technique stops working.

Sign in to see when this approach fails.

2–4 min · Everything important, quickly.

Sign in to see the quick scan of the deep dive.
MethodWhat it quantizesFailure mode it fightsTypical bit width
GPTQWeights onlyCumulative layer reconstruction errorINT4 / INT3
AWQWeights onlyDestruction of salient weight channelsINT4
SmoothQuantWeights and activationsActivation outliers blocking INT8INT8 (W8A8)
FP8 cast-and-goWeights (and often activations)Narrow dynamic range of INT8FP8 (E4M3)

Real products, models, and research that use this idea.

  • vLLM and SGLang in 2026 ship AWQ and GPTQ INT4 kernels as a default way to serve Llama 3.1 70B on a single 80 GB GPU.
  • TensorRT-LLM uses SmoothQuant to enable W8A8 INT8 serving on NVIDIA H100 tensor cores for latency-sensitive deployments.
Sign in to see more production examples.

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy does AWQ scale activations down when it scales salient weight channels up?
A

The product of weights and activations must stay numerically equivalent. AWQ multiplies a salient weight channel by a factor and divides the matching activation channel by the same factor, so the matmul output is unchanged while the weight values now occupy more of the INT4 range.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Treating all four as interchangeable INT4 tricks. They differ on what they quantize (weights only vs weights and activations) and which failure mode they target (salient channels, layer error, or activation outliers).

Sign in to see all red flags and common mistakes.

60 second bullets to scan on the way to the call.

  • Which methods are weight only versus weights and activations

  • How GPTQ uses the Hessian to minimize reconstruction error

Sign in to unlock the revision sheet.

Primary sources. Browse if you want the original framing.

Similar questions

Same topic, related formats. Practice these next.

4 curated
Next question
What is the KV cache in transformer inference?
Flashcard·Easy