Qwen2.5-1.5B-Instruct on llama.cpp (Metal), measuring the production KPIs (TTFT, TPOT, ITL, tokens/sec) across quantization, context length, prefix caching and model load time.