Scholar Feed

LLM quantization

AffineQuant

AffineQuant: Affine Transformation Quantization for Large Language Models

Superseded baseline#26 of 80 most-superseded · first seen Mar 19, 2024

Superseded — cited as a baseline and beaten by newer methods

2 papers critique it · 1 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites AffineQuant as a baseline.

However, due to the significant overhead of full-size matrix multiplication, AffineQuant can only apply affine transformations to a small fraction of linear layers.
— InfoQuant: Shaping Activation Distributions for Low-Bit LLM Quantization
“While affine transformations theoretically offer greater flexibility than rotation transformations for handling outliers, the original AffineQuant approach has practical limitations. It learns a full transformation matrix that can only be applied to output projection layers for weight-activation quantization, where it merges with preceding linear layers to avoid overhead. Other layers must use per-channel scaling, limiting the method's broader applicability across model architectures.”
— Adaptive Layer-Wise Transformations for Post-Training Quantization of Large Language Models

Beaten on benchmarks

Head-to-head results where a newer method reports beating AffineQuant. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.