inference engine

i use llms every day, and i realised i couldn't explain what happens between my prompt and the first word that comes back. that bugged me.

right now it's mostly reading and notes: tokenizers, attention, kv caches, quantization. once it clicks, i'm writing an engine in rust that loads open weights and runs them on my own machine, no python anywhere.

next up is getting a small model to spit out its first token. then making it fast.