Warmup
Also known as warmup phase, cache priming
In plain English
Warmup is the priming pass before measurement starts, so the run is scored on a system in steady state rather than one with an empty cache.
Technical definition
Warmup is the phase before the profiling window in which a benchmark primes caches and reaches steady state, and whose requests are excluded from reported results.
Engineering details
A cold system misrepresents production twice over: the cache is empty, and every session starting at turn zero creates a synchronized burst that no real deployment sees. AgentX therefore starts each conversation at a seeded point partway through its history, replays the requests needed to reconstruct that state, then advances each replay lane further before measurement begins.
Why it matters
Where the measurement window starts changes the result. Include the priming pass and prefix reuse looks worse than production; skip priming entirely and cache-dependent recipes are scored on a state they would never serve from.
How to read it in InferenceX
The AgentX point view separates the two phases, so telemetry can be inspected for either. Warmup requests are capped at a single output token, which is why their output length is about one and their interactivity and decode series are blank: one token has no inter-token latency.
Source material
See the concept in real benchmarks
AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?
$3 Million USD dataset open sourced, 1 Mil+ Context Length, Multiturn, Sub Agents 95%+ KVCache HitRate, GB300 NVL72, MI355X, B200
Agentic Benchmark for LLM Inference: Metrics and Methodology
How an agent benchmark replays long-context, multi-turn workloads to measure latency, throughput, cache behavior, and serving cost
A Brief Overview of Agentic Workloads
Multi-turn sessions, long contexts, and near-total prefix reuse make agentic inference a systems problem, and change what a benchmark has to measure