Superfluid's serving layer introduces no additional overhead: under single-request and eight-concurrent-request conditions, the tokens/s of llama.cpp called through Superfluid is on par with using llama-server directly.