Resin

About

Resin is an experimental library of building blocks for LLM inference engines. It provides a headless Rust engine with a scheduler, KV cache and metrics server, and the user writes the forward pass as model-specific kernels in Python. The authors say they built engines for two models on Blackwell GPUs in 12 hours and that results come within 5 to 10 percent of other engines. The README warns it is not production quality.