SALT: CELF-Based Sentence-Level Compression for KV Cache Retrieval

No_Sky9786 · reddit · 2026-08-19

The author introduces SALT, a structure using the CELF algorithm to retrieve only important sentences from the KV Cache. It shrinks long documents to a fixed size before sending them to the LLM, retaining the most informative sentences. Compatible with any model, it produces shorter prompts and cuts compute, memory, and latency. The project is open-source on GitHub. The author is seeking help to implement a dynamic budget adjustment method instead of a fixed 20-25% retrieval rate to optimize GPU prefill memory usage at scale.

Original post →

More from Infra

Infra channel →