LLM prefix KV cache sharing

Design the prompt/KV-cache layer for an LLM-serving system where most requests share a large, mostly-static prefix (a long system prompt + retrieved context) and differ only in a short user suffix. At 5k concurrent requests, recomputing the prefix's attention KV every time wastes most of the GPU. Walk through how you cache and reuse the prefix KV across requests, how you manage GPU memory as a finite resource shared by all in-flight sequences, and the correctness traps in reusing cached KV.