Token By Token

Episode 04 · Planned

Batching, scheduling and mixed traffic

Study chunked prefill and scheduler behavior under controlled long/short traffic mixes.

Hold the workload fixed while changing chunked-prefill support or configuration, chunk size and token budget one variable at a time. Then vary the long/short traffic mix. Continuous batching is part of the scheduler; do not assume every engine exposes a valid off mode. A static-batch reference is a separate workload if the runtime supports it.

Evidence: verified effective flags, running/waiting requests, scheduled tokens per iteration when exposed, queue time, TTFT and decode gaps by request class, maximum wait, starvation indicators and KV time series.

Read this experiment in the roadmap →