The common optimization trick for speeding up transformer inference is KV caching 1 2. This technique is so prominent that huggingface library has use_cache flag is enabled by default 6. A few days ago, I read an awesome blog post on GPT in 60 Lines of NumPy.…