Show HN: UL-SMF – Open-source linear-complexity ~300x KV-cache compression

 (github.com)

13 points | by liventruth 22 hours ago

2 comments

  • colingauvin 11 hours ago
    This should be trivially demonstrable if it actually works. Cosine similarity is not a good metric, however, for attention compression.
  • liventruth 22 hours ago
    [flagged]