vLLM is probably the key difference there… its scheduler is built around batching/concurrency, while this setup is heavily optimized llama.cpp for single-stream latency