[Feature] metrics support

[CUHKSZzxy]

Objective

Align with vLLM v1 metrics system and beyond. Here are several key alignments

• Monotonic Timestamps:
  -- Uses time.perf_counter() for interval calculations (avoids clock drift issues).
• Metric Types:
  -- Gauges: Active requests, cache usage, etc
  -- Counters: Token totals, request success / failure counts, etc
  -- Histograms: TTFT (Time-To-First-Token), TPOT (Inter-Token Latency), end-to-end latency, etc
• Metrics Publishing:
  -- CLI logging
  -- Prometheus & Grafana

We only record critical timestamps and events inside the engine process without further processing. Heavy-weight metrics calculations or publishing are separated from the main loop to minimize overhead.

For convenient Grafana visualization and usage, we align with SGLang.

TODO

• Refactor: After MP engine, things become different ... the global singleton context will be local for each process. Carrying the information from the engine to the async engine seems the most convenient and less error-prone way to do it; otherwise, we may perform IPC frequently.
• Refactor: 1. Avoid parameter passing (singleton context), 2. Reduce computation overheads (high CPU overheads, but can be solved with MP engine)
• Refactor: Decouple prometheus_client, only install / import when needed
• Update: Add user guide
• Refactor: Reduce messy parameters, pack things into a class
• Feature: Grafana visualization
• Feature: Expert information collections (deferred in another PR)
• Refactor: Minimize the modifications to async engine generate() and engine _async_loop_main()
• Fix: Use time.perf_counter()

Usage

Start the server with --enable-metrics

lmdeploy serve api_server Qwen/Qwen2.5-7B-Instruct --enable-metrics

• Metrics Publishing - Logging
  With --enable-metrics, key metrics (e.g., finished / unfinished / running / waiting requests, token throughputs, cache usage) are printed to the terminal every 10 seconds.
  

• Metrics Publishing - Prometheus & Grafana
  -- Raw Metrics
  Access the raw Prometheus metrics via http://localhost:23333/metrics/ .
  You can also curl the metrics endpoint curl http:///localhost:23333/metrics/ to view raw Prometheus results. No extra setups are required for this step.
  
  -- Prometheus Panel
  Access the Prometheus panel via http://localhost:9090 (9090 is the current default port for the Prometheus panel). You need extra setups to access the Prometheus panel; please check the user guide for details.
  
  -- Grafana Panel
  Access the Grafana panel via http://localhost:3000 (3000 is the current default port for the Grafana panel). You need extra setups to access the Grafana panel; please check the user guide for details.
  

Request Timeline

The following diagram depicts how we define and calculate time intervals during the request lifecycle, which adheres to vLLM.

Performance Impacts

• Conclusion

Tested with Qwen2.5-0.5B / Qwen2.5-7B / Qwen2.5-32B, no obvious performance impacts. (Requires #3627)

Check the following tables for output throughput details. We conducted tests using 1,000 prompts, with input length 1k and output length 1k. Each model was tested three times to reduce the impact of performance fluctuations.

• QWen2.5-0.5B, TP1

W/O metrics (tokens/s)	W metrics (tokens/s)
20387	20555
20341	20877
20746	20771

---

• QWen2.5-7B, TP1

W/O metrics (tokens/s)	W metrics (tokens/s)
8836	8721
8780	8736
8800	8723

---

• QWen2.5-32B, TP2

W/O metrics (tokens/s)	W metrics (tokens/s)
3019	3160
3167	3165
3189	3173

Related Issues & PR

• [Feature] Metrics Endpoint #2638
• [Feature] Response Metrics #2673
• Add metrics endpoint #1423

the design is changed

[RunningLeon]

can we config all dp server urls in here and show data in grafana board?

[RunningLeon]

LGTM

[grimoire]

LGTM