LLM Serving Capacity Planning

Planning usable context and concurrency on a fixed GPU: how to find the real memory bottleneck before optimizing, and why the default config often wins.