verifiedhigh · 0.88updated 2026-10-03

KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache (Yuan et al., LMSYS)

paper

Fields

sources/qlab-kivi-paper

URL

Publisher

arXiv

Kind

paper

Published

2024-02-05

Accessed

2026-10-03

notes

v2. Per-channel keys, per-token values; the source for the KV key/value bit asymmetry.

v2. Per-channel keys, per-token values; the source for the KV key/value bit asymmetry.

This record references no other records.