跳到主要内容Skip to main content

量化方法Quantization Methods (GPTQ, AWQ)

Quantization Methods (GPTQ, AWQ)

  • 硬件与部署Hardware & deployment
  • 第五站 · 跑起来Station 5 · Running It
解决什么问题What problem it solves

把模型压小不是只有一种做法,不同方法效果不同。Shrinking a model has more than one method, and the methods differ in results.

这是什么What it is

GPTQ 和 AWQ 是最常见的两族训练后量化方法。

它们都靠少量样本校准,决定哪些数值该保留精度。AWQ 按激活值的重要性挑选保留的权重。

GPTQ and AWQ are the two most common families of post-training quantisation.

Both calibrate with a small sample to decide which values keep their precision; AWQ picks the weights to keep by how much each matters to activations.

为什么需要知道Why it matters

同一个模型的不同量化档,可能出自不同方法。同名 4-bit 文件,质量未必一样。Different quantisation tiers of the same model may come from different methods. Two files both labelled 4-bit need not be equal in quality.

别名Chinese aliases

GPTQ · AWQ

出处Sources
  • GPTQ 论文GPTQ paper ↗
  • AWQ 论文AWQ paper ↗