CachyLLama caches KV state on SSD and cuts 15,700-token prompts to under 1s

UsedMorning9886 · reddit · 2026-07-25

CachyLLama caches KV state on SSD to speed up local agent workflows

CachyLLama is a llama.cpp fork built for local agentic coding setups such as Aider and Claude Code, where prompt evaluation can dominate runtime.

What it changes

Benchmarks on AMD Ryzen 7840U / 780M

The project does not speed up token generation itself; it mainly removes repeated prompt-processing overhead, which is exactly the bottleneck many local agent harnesses hit.

Related event: CachyLLama Persists KV Cache to Cut Local Agent Overhead(2 posts)→

Original post →

More from coding & agent

coding & agent channel →