[PyTorch] Best practices for optimizing inference speed and memory footprint with custom PyTorch models?

Part of

Cross-Architecture Performance Modeling Trends