Hi Product Hunt, Profile's engineer here! Most GPUs serving LLMs deliver a fraction of what the hardware can do, and the gap hides behind a handful of server flags. Finding the right flags is usually days of trial and error. Profile replaces the guessing. It computes the fastest your GPU can physically serve your model, watches your live server, and names the one thing slowing it down, with the exact flag to change. You apply it, Profile measures the difference. On my RTX 5090 that loop took a 30B model from 81 to 421 tokens/s in 4 iterations, about 30 minutes. Cost per million tokens fell 81%. On an H100 running Qwen: 1.9x, cost down 48%. Both runs are on video, including the step that made things worse; Profile labeled it "worse" and the next fix recovered it. A tool that only reports wins can't be trusted when it reports one. Free, open source (Apache 2.0), works with vLLM today, more engines next. Ask me anything, and tell me what it finds on your serve

Profile

Physics grounded, & cost aware, optimization loop for vLLM

Physics grounded, & cost aware, optimization loop for vLLM