Bringing Flame Serving Close to Standalone vLLM

A technical report compares Flame, a Python service embedding vLLM, against standalone vLLM serving Llama 3.1 8B on two B200 GPUs. After adapter tuning, Flame's median throughput was within 1% of standalone vLLM across five paired trials, for both no-reuse and prefix-reuse bursts.