Between llama.cpp, vLLM, Ollama, and the various GUIs, the local tooling landscape is crowded. I'm setting up a team workstation setup for prototyping before cloud spend. What's your current local stack for running 7B-70B models, and what would you avoid?