Prefill stage latency varies wildly with KV-cache layout making baseline modeling noisy

Prefill execution times fluctuate drastically with long-tail distribution even for identical input lengths due to KV-cache hit rate variations from PagedAttention and RadixAttention optimizations, making physical baseline modeling prone to noise

Prioritize Decode stage latency monitoring for anomaly detection rather than Prefill. Accept Prefill variability as normal operational characteristic when PagedAttention/RadixAttention optimizations are enabled. Focus user-experience monitoring on Time-Between-Tokens rather than Time-to-First-Token.