Efficient training and inference is table stakes for LLMs these days, and these two algorithmic efficiency techniques work really well for reducing LLM latency as well as memory usage, while retaining the model performance. Feel free to give it a read, and drop a note if I missed something.