Walk me through how you would optimize the inference latency and cost of a large language model serving millions of requests per day.
