FasterGPU, for the GPUs already in the rack

Field notes

The site frames inference tuning as a sign-off problem, not merely a speed claim: it measures the existing deployment, agrees on a target, and supplies before-and-after figures that the client can rerun. Its examples name the hardware, model, precision, and serving stack, including a Qwen2.5-7B RTX 4090 FP16/vLLM case moving from 63.0 to 90.5 tokens per second. The homepage also reports a 55.5% average FP8/FP16 prefix agreement across six prompts, with five diverging.