TurboPrefill by Trykhlieb is now also available for ik_llama.cpp by Ivan Kawrakow.

Measured prefill acceleration at 16,373 context tokens:

2× RTX 1080 Ti: +60%

4× RTX 3090: +90%

5× RTX 5060 Ti: +90%

8× RTX 5060 Ti: +120%

10× P104-100 Pascal: +350%

RFC / PoC discussion: https://github.com/ggml-org/llama.cpp/pull/24219