OSS
UL-SMF Posts an Open-Source Linear-Complexity KV-Cache Compression Claiming ~300x, 21 Hours Old on Show HN
liventruth/UL-SMF-Cache-Compression appeared on Show HN 21 hours ago claiming roughly 300x KV-cache compression at linear complexity, currently at 13 points and 1 comment. The claim is extraordinary relative to published KV-compression work, the engagement is near zero, and there is no third-party reproduction — so this is a single-source flag, not a result. It is worth watching precisely because it is adjacent to what oMLX is doing with SSD-tiered caching: if long-context serving on modest hardware is the bottleneck, cache compression and cache tiering are the two live attacks on it.
Source
↳ Follow the thread