vLLM KV Cache Math: Why My 24GB GPU Served Only 4 Requests at Once

vLLM KV cache sizing explained: bytes per token, why a 24GB GPU runs 4 RAG requests at once, and the flags that raise concurrency.