Prefix/KV-cache reuse

Serving optimization that stores and reuses the attention KV cache for shared token prefixes across requests, cutting redundant prefill compute and latency.

grounded in: the local vLLM inference stack (vllm concept; spark.sh / router_stats.py serving GB10 nodes) — prefix/KV-cache reuse is a core, durable vLLM serving capability

Connected concepts

vLLM / PagedAttention, Distributed inference, Speculative decoding, Test-time compute