AI product
AdmitPerf is an open-source admission-control library and benchmark harness for self-hosted LLM inference. As an ingress library, it decides whether to admit, defer, or reject each request before it reaches an inference engine such as vLLM or SGLang. Its benchmarking mode runs admission policies against a real engine on a real GPU and records seven disclosure items by default, so the evaluated policy corresponds to the one deployed in production.
1 use taken from transcripts — each links to the moment in the video.
The authors' own open-source library and benchmark harness: in production it is the ingress library deciding admit/defer/reject for every request in front of vLLM or SGLang; in benchmarking it runs any policy against a real engine on a real GPU and records the seven disclosure items by default, so the benchmarked policy is exactly the deployed one
1 in the library.