fix metrics

[CUHKSZzxy]

Fine-grained metrics calculation:

Dataflow: client --> API server --> Engine core

API server request states (axis view):
|<──────────────────────────────── total ────────────────────────────────>|
|<──────────── completed ─────────────>|<────── uncompleted ─────────────>|
|<─ success ─>|<──────── fail ────────>|<─ routed ─>|<───── waiting ─────>|
              |<cancel>|<abort>|<error>|

Engine core request states (axis view):
|<────────────────── routed ──────────────────>|
|<───── running ──────>|<────── waiting ──────>|

Now it looks like:

[2026-03-13 14:54:59 DP0] Avg thr (in/out): 1445.8 / 68.8 tokens/s, Server (succeeded/failed/routed/waiting): 102 / 100 / 0 / 0, Engine (running/waiting): 0 / 0, KV cache: 17.5%,

[Copilot]

Pull request overview

This PR refactors the metrics system to provide fine-grained request state tracking. Instead of a single "completed" counter, requests are now categorized as succeeded, cancelled, or aborted. The num_api_waiting_reqs is now a computed property derived from total - completed - routed rather than being manually calculated in multiple places.

Changes:

• Replaces num_completed_reqs with num_succeeded_reqs, num_cancelled_reqs, and num_aborted_reqs in SchedulerStats, with computed properties for derived metrics (num_failed_reqs, num_completed_reqs, num_uncompleted_reqs, num_api_waiting_reqs).
• Updates MetricsProcessor to expose increase_succeeded_requests, increase_cancelled_requests, and increase_aborted_requests methods, and updates call sites in async_engine.py.
• Updates Prometheus gauges and log messages to report the new fine-grained metrics.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
lmdeploy/metrics/stats.py	Replaces num_completed_reqs field with fine-grained counters and computed properties
lmdeploy/metrics/metrics_processor.py	Adds increase_succeeded_requests, increase_cancelled_requests, increase_aborted_requests methods
lmdeploy/serve/core/async_engine.py	Updates call sites: cancelled in exception handler, aborted for pre-start abort, succeeded after generator loop
lmdeploy/metrics/loggers.py	Updates log format and Prometheus gauges for new metric names

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.