LLM Serving Benchmark/Research
Controlled benchmark of vLLM vs llama.cpp under identical workloads, measuring TTFT, ITL, throughput, goodput, memory, and quality across quantization schemes.
Backends run as separate server processes under deterministic prompt generation, with client-side timing on a monotonic clock, warmup passes discarded, and unavailable cells recorded as skipped with a reason.
Tell me what you're building, the constraints you're working with, and where it breaks. I reply within a day.