In real-time interaction models like Moshi, MiniCPM-o, and Qwen-Omni, serving becomes a periodic real-time task: each frame a session ingests streaming audio and must respond by a recurring wall-clock deadline, while its KV cache grows monotonically and stays pinned throughout the conversation. This regime hides a dangerous failure mode.
In a real full-duplex stack, sustained load does not gracefully degrade service: it falls off a cliff, jumping from milliseconds per frame to a stalled engine when accumulated session state exhausts the KV pool. The collapse is metastable—identical five-minute runs may collapse or survive due to run-to-run variance—and silent: latency and deadline-miss metrics appear healthy throughout.
We show one move restores both stability and observability: bound each session's resident state, and latency starts telling the truth. Metronome's in-engine KV window eliminates the collapse (0/20 vs. 14/20 runs across two batches) and turns per-frame latency into a monotone load signal, on which an online admission controller discovers the schedulable concurrency; without the window, the identical controller over-admits into the wall.
A first-order model predicts the collapse time within a few percent on the headline model, and a quality probe validates the bound's design by ablation: the window alone is quality-free in turn-based decoding, and its few pinned attention-sink tokens are what keep free-running generation healthy. Everything is measured end-to-end on real audio, across four interaction models on one GPU.
Blogger's Review: Metronome addresses potential collapse issues in real-time interaction model serving by bounding the use of KV caches, demonstrating high stability and observability. This technology not only enhances system reliability but also provides an effective solution for real-time processing, making it worthy of attention and study.