# 请问我在使用vllm 推理qwen3-vl的时候多次请求的过程中比如发起5个请求后，发送的内容是一致的，前面4个请求相对较快返回，最后一个请求，也在running没有处于waiting中，但是等待的结果却比他们慢了很多，比如前面四个可能40s能返回，最后一个需要6分钟

**URL:** <https://discuss.vllm.ai/t/vllm-qwen3-vl-5-4-running-waiting-40s-6/2104>\
**Category:** General\
**Created:** [December 11, 2025, 6:00am UTC](https://discuss.vllm.ai/t/vllm-qwen3-vl-5-4-running-waiting-40s-6/2104 "2025-12-11T06:00:25Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![Squidward-J](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/squidward-j/32/995_2.png) [@Squidward-J](https://discuss.vllm.ai/u/Squidward-J)\
**Post date:** [December 11, 2025, 6:00am UTC](https://discuss.vllm.ai/t/vllm-qwen3-vl-5-4-running-waiting-40s-6/2104/1 "2025-12-11T06:00:25Z")

</div>

![0f105ec4-e9be-4d44-abb0-09a8c04e6d86](https://canada1.discourse-cdn.com/flex036/uploads/vllm/original/1X/e51808d531ca6316fa474fae2b148d42867e6a66.png)

---

<div class="post-metadata">

**Author:** ![RunLLM](https://yyz2.discourse-cdn.com/flex036/user_avatar/discuss.vllm.ai/runllm/32/233_2.png) [@RunLLM](https://discuss.vllm.ai/u/RunLLM)\
**Post date:** [December 11, 2025, 6:01am UTC](https://discuss.vllm.ai/t/vllm-qwen3-vl-5-4-running-waiting-40s-6/2104/2 "2025-12-11T06:01:02Z")

</div>

The dashboard visualizes key vLLM production metrics: E2E Request Latency (P99, P95, P90, Avg), Token Throughput (prompt/generation tokens/sec), Time Per Output Token Latency (various percentiles), and Scheduler State (running/waiting requests). The spike at 13:50 in both latency and throughput, along with a peak in running/waiting tasks, suggests a temporary load increase or resource contention, followed by recovery and stabilization—this is typical for systems under bursty traffic or resource constraints, as described in [vLLM metrics documentation](https://docs.vllm.ai/en/latest/design/metrics/) and [Grafana dashboard examples](https://docs.vllm.ai/en/latest/examples/online_serving/dashboards/).

These metrics help diagnose performance bottlenecks, monitor SLOs, and understand how request load impacts latency and throughput. Would you like a detailed explanation of how to interpret or act on these metrics?

Sources:

- [vLLM Metrics Documentation](https://docs.vllm.ai/en/latest/design/metrics/)

Hint: Mention @RunLLM in the post for followups.