2 ms·LLM Efficiency: From Hyperscale Optimizations to Universal Deployability1 points by PaulHoule 11mo ago