[Feature] metrics support #3534
Open
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Conflicts: lmdeploy/messages.py lmdeploy/pytorch/engine/engine.py lmdeploy/pytorch/engine/engine_instance.py lmdeploy/pytorch/messages.py lmdeploy/pytorch/paging/scheduler.py
Conflicts: lmdeploy/serve/openai/api_server.py
lvhan028
reviewed
May 29, 2025
lmdeploy/serve/async_engine.py
Outdated
| @@ -302,6 +303,21 @@ def __init__(self, | |||
| self.internal_thread = _EventLoopThread(daemon=True) | |||
| self.limiter: asyncio.Semaphore = None | |||
|
|
|||
| # build status loggers | |||
| # independent set for each DP rank, since monototic time differs for each process | |||
| # each set contains one cli logger and one prometheus logger | |||
There was a problem hiding this comment.
It might be cleaner to move the new code into a separate function
lvhan028
reviewed
May 29, 2025
| # actual running requests | ||
| num_running_reqs = self.scheduler.num_locked() | ||
| # waiting to be scheduled or have been scheduled but not yet started execution | ||
| num_waiting_reqs = self.scheduler.num_waiting() + self.scheduler.num_running() |
There was a problem hiding this comment.
Does self.sched
8000
uler.num_running() refer to scheduled but not yet started?
There was a problem hiding this comment.
yes, discussed and confirmed with yaoq.
lvhan028
reviewed
May 29, 2025
| @@ -997,6 +1062,16 @@ def __send_resps(step_outputs: List[InferOutput]): | |||
| await self._await_forward_event(forward_event) | |||
| __send_resps(resps) | |||
|
|
|||
| async def _async_log_stats_task(self, log_que: asyncio.Queue): | |||
|
|
|||
| while True: | |||
There was a problem hiding this comment.
Can this task be terminated normally?
grimoire
reviewed
May 30, 2025
| @@ -1149,14 +1227,19 @@ async def async_loop(self): | |||
| forward_event, has_runable_event 9E81 ), | |||
| name='MainLoopPreprocessMessage') | |||
|
|
|||
| # log task | |||
| logger.info('Starting async task MainLoopLogStats.') | |||
There was a problem hiding this comment.
do not create this task if metrics is disabled.
|
May merge the main branch to resolve lint errors |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Objective
Align with vLLM v1 metrics system and beyond. Here are several key alignments
-- Uses
time.perf_counter()for interval calculations (avoids clock drift issues).-- 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
-- CLI logging
-- Prometheus & Grafana
We only record critical timestamps and events during the main loop and scheduling without further processing. Heavy-weight metrics calculations or metrics publishing are put inside separate coroutines to reduce the main engine loop overhead.
TODO
time.perf_counter()generate()or engine_async_loop_main()Expert information collections (deferred in another PR)Usage
Start the server with
--enable-metricsMetrics Publishing - Logging
With
--enable-metrics, key metrics (e.g., running / waiting requests, cache usage, token throughput) are printed to the terminal every 5 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.-- Prometheus Panel (WIP, user guide to be added)
Access the Prometheus panel via http://localhost:9090 (
9090is the current default port for Prometheus panel)-- Grafana Panel (WIP, user guide to be added)
Access the Grafana panel via http://localhost:3000 (
3000is the current default port for the Grafana panel)Performance Impacts
Conclusion:
You may check the following figures for details. Benchmark settings: 1000 prompts, input len 1000, output len 1000.
Qwen-2.5-7B (TP1), without the metrics.
Qwen-2.5-7B (TP1), with the metrics.
Qwen-2.5-0.5B (TP1), without the metrics.
Qwen-2.5-0.5B (TP1), with the metrics.
Related Issues & PR
Issue 2638, Issue 2673, PR1423