The authors trained a selector that assigns each transformer layer a cache configuration under a specified memory budget.
Adaptive per-layer choices shrink LLM caches without sacrificing accuracy
KV-Kaizen learns how to combine cache sharing, reduced precision and rank truncation for each layer and input context.
Big Tech
Joao Monteiro · Louis Béthune · Anastasiia Filippova · Sonia Laguna · David Grangier · Marco Cuturi
Apple
Research Digest··2 min read
Monteiro et al.
Why this paper
From Apple
In one line
A learned per-layer selector compresses KV caches along depth, precision, and rank axes without eviction, preserving accuracy at high compression.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§