The standard guidelines for building large language models (LLMs) optimize only for training costs and ignore inference costs. This poses a challenge for real-world applications that use ...