I've been using a remote ollama server with a local jupyter notebook. The langchain configuration allows me to specify the ollama host. So I can develop locally with remote models. I guess I still don't see the difference. Does lepton decouple the HTTP server from the model backend?