A simple fix for LLM tail latency