Scholar Feed

KV-cache compression

QuaRot

QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs

Superseded baseline#46 of 234 most-superseded · first seen Mar 30, 2024

Superseded — cited as a baseline and beaten by newer methods

2 papers critique it · 2 beat it on benchmarks

What papers say

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

However, a random rotation is still data-oblivious. It can smooth activation ranges, but it does not know which directions are important to attention. At INT2, this distinction matters: only four quantization levels are available, so the error should be pushed into directions that the model reads less strongly.
— OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization
“While this approach has shown effectiveness at moderate precision levels, such as a 4-bit KV cache, its applicability under more aggressive quantization settings remains unexplored.”
— KVLinC : KV Cache Quantization with Hadamard Rotation and Linear Correction

Beaten on benchmarks

Head-to-head results where a newer method reports beating QuaRot. 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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.