Why it matters

GPTQ enables 4-bit LLMs. Understanding shapes quantization strategy.

Advertisement

The architecture

Per-layer quantization.

Hessian approximation.

Cholesky-based error correction.

GPTQ flowPer-layerfp16 weightsCompute HessianactivationsQuantize witherror correctionAutoGPTQ + GPTQ-for-LLaMA libraries; W4A16 typical
GPTQ.
Advertisement

How it works end to end

Frantar 2023.

Requires calibration data.

W4A16 (4-bit weights, 16-bit activations).

Group-wise scaling.