KV-cache compression

Mooncake

Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving

Superseded baseline#58 of 234 most-superseded · first seen Jun 24, 2024

Superseded — cited as a baseline and beaten by newer methods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 Mooncake as a baseline.

Mooncake uses reactive placement---blocks are placed based on current utilization---while our Bayesian predictor proactively positions blocks based on predicted reuse probability, achieving 70--84% hit rates vs. Mooncake's reported 65--80%
Predictive Multi-Tier Memory Management for KV Cache in Large-Scale GPU Inference
their approaches are restricted to sharing in the layer or text segment within adjacent layers or the same LLM, limiting the broader applicability
SemShareKV: Efficient KVCache Sharing for Semantically Similar Prompts via Token-Level LSH Matching

Beaten on benchmarks

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Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.