In this second article in our short series on SLM optimization techniques we focus on the reuse of the prompt prefix with a key-value cache.