The Slowest Kid Problem: How a Super Captain Solves MoE's Biggest Headache
Why a mixture-of-experts layer runs at the speed of its busiest expert, how to measure it, and when predictive prefetching actually helps.
Mixture of Experts02
Subject · 2 posts
Why a mixture-of-experts layer runs at the speed of its busiest expert, how to measure it, and when predictive prefetching actually helps.
Mixture of Experts02
Big-O counts steps and assumes every step costs the same. It does not. A cache miss costs a hundred steps, and that is why a linked list loses to an array.
Problem Solving From Zero21