[V1][Metrics] Add initial Prometheus logger (#12416)

Signed-off-by: Mark McLoughlin <markmc@redhat.com>
This commit is contained in:
Mark McLoughlin
2025-01-27 12:26:28 -05:00
committed by GitHub
parent 103bd17ac5
commit 01ba927040
3 changed files with 78 additions and 10 deletions
+7 -4
View File
@@ -24,7 +24,8 @@ from vllm.v1.engine.core_client import EngineCoreClient
from vllm.v1.engine.output_processor import OutputProcessor
from vllm.v1.engine.processor import Processor
from vllm.v1.executor.abstract import Executor
from vllm.v1.metrics.loggers import LoggingStatLogger, StatLoggerBase
from vllm.v1.metrics.loggers import (LoggingStatLogger, PrometheusStatLogger,
StatLoggerBase)
from vllm.v1.metrics.stats import IterationStats, SchedulerStats
logger = init_logger(__name__)
@@ -46,13 +47,15 @@ class AsyncLLM(EngineClient):
assert start_engine_loop
self.model_config = vllm_config.model_config
self.log_requests = log_requests
self.log_stats = log_stats
self.stat_loggers: List[StatLoggerBase] = [
LoggingStatLogger(),
# TODO(rob): PrometheusStatLogger(),
PrometheusStatLogger(labels=dict(
model_name=self.model_config.served_model_name)),
]
self.model_config = vllm_config.model_config
# Tokenizer (+ ensure liveness if running in another process).
self.tokenizer = init_tokenizer_from_configs(
@@ -272,7 +275,7 @@ class AsyncLLM(EngineClient):
# 4) Logging.
# TODO(rob): make into a coroutine and launch it in
# background thread once we add Prometheus.
# background thread once Prometheus overhead is non-trivial.
assert iteration_stats is not None
self._log_stats(
scheduler_stats=outputs.scheduler_stats,
+36
View File
@@ -1,5 +1,8 @@
import time
from abc import ABC, abstractmethod
from typing import Dict
import prometheus_client
from vllm.logger import init_logger
from vllm.v1.metrics.stats import SchedulerStats
@@ -36,3 +39,36 @@ class LoggingStatLogger(StatLoggerBase):
scheduler_stats.num_running_reqs,
scheduler_stats.num_waiting_reqs,
)
class PrometheusStatLogger(StatLoggerBase):
def __init__(self, labels: Dict[str, str]):
self.labels = labels
labelnames = self.labels.keys()
labelvalues = self.labels.values()
self._unregister_vllm_metrics()
self.gauge_scheduler_running = prometheus_client.Gauge(
name="vllm:num_requests_running",
documentation="Number of requests in model execution batches.",
labelnames=labelnames).labels(*labelvalues)
self.gauge_scheduler_waiting = prometheus_client.Gauge(
name="vllm:num_requests_waiting",
documentation="Number of requests waiting to be processed.",
labelnames=labelnames).labels(*labelvalues)
def log(self, scheduler_stats: SchedulerStats):
"""Log to prometheus."""
self.gauge_scheduler_running.set(scheduler_stats.num_running_reqs)
self.gauge_scheduler_waiting.set(scheduler_stats.num_waiting_reqs)
@staticmethod
def _unregister_vllm_metrics():
# Unregister any existing vLLM collectors (for CI/CD
for collector in list(prometheus_client.REGISTRY._collector_to_names):
if hasattr(collector, "_name") and "vllm" in collector._name:
prometheus_client.REGISTRY.unregister(collector)