Cache reuse

An advanced technique for reducing the
prompt-processing time by "shifting" chunks
of the previous context to new positions.

                    # prompt 0 (cached)

                    AAABCCCCDDDEEFGGHHHIIIxxx
                    
                    # prompt 1 (reuse from prompt 0)

                    AAACCCCEEGGIIIyyy

req:

RoPE encoding

uses:

Partial context updates, Thinking mode

> llama-server --cache-reuse 256 [...]