prefix caching in llama.cpp with PyTorch

[bytesupply76]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Kimi and compare from there. Not glamorous advice, but repeatable baselines save a lot of time.