The same ten requests are already queued in the premium band. Numbered markers identify requests; dispatch leaves right. These are illustrative inputs, not a benchmark or model-execution timeline. Time zero in the animation is the queued snapshot at 10 seconds in the input table.

FCFS uses enqueue time. EDF uses enqueue time + effective TTL. SLO Deadline uses received time + the request's TTFT target header. The earliest deadline sorts first; those deadlines are absolute offsets from a shared origin.

Round Robin takes turns across active queues. Global Strict compares queue heads using the ordering policy. Program-Aware is shown with its LAS strategy and default service/wait weights (0.8/0.2). Prior weighted-token service is A: 900, B: 100, C: 400. No requests finish during this dispatch example, so there are no new service charges; common decay preserves service ratios. These inputs make B, then C, then A win even with the waiting-time term. Turn-priority is another Program-Aware strategy, outside this example.

Source: llm-d-router d4b8afd3. Fairness operates within the selected priority band. Dispatch is still subject to admission and endpoint selection; it does not establish equal throughput or an SLO guarantee.