SGLang vs vLLM:优先级调度、限流、淘汰策略对比
一、优先级调度
| 维度 | SGLang | vLLM |
|---|---|---|
| 默认策略 | FCFS(First Come First Serve) | FCFS |
| 优先级模式 | --enable-priority-scheduling | scheduling="priority" |
| 优先级方向 | 默认高数值=高优先级;schedule_low_priority_values_first可反转 | 低数值=高优先级(min-heap) |
| 排序方式 | (priority * sign, wait_queue_entry_time) | (priority, arrival_time, request_id) |