AirLLM 70B inference with single 4GB GPU

Part of

Cross-Architecture Performance Modeling Trends