A complete Llama2 inference engine that fits in 1356 bytes of x86 assembly / hacker news