A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 20:45 UTC
[Metrics] Enable sleep state metric outside of dev mode (#27867)
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Mark McLoughlin committed
Nov 4, 2025 at 04:35 UTC
380ba6816d4646be99d9b6d207ba7bc7fce8290e
1 file changed
+23
-27
vllm/v1/metrics/loggers.py
+23
-27
@@ -9,7 +9,6 @@ from typing import TypeAlias
9
10
from prometheus_client import Counter, Gauge, Histogram
11
12
-import vllm.envs as envs
12
from vllm.config import SupportsMetricsInfo, VllmConfig
13
from vllm.distributed.kv_transfer.kv_connector.v1.metrics import (
14
KVConnectorLogging,
@@ -395,32 +394,32 @@ class PrometheusStatLogger(AggregateStatLoggerBase):
394
self.gauge_scheduler_waiting = make_per_engine(
395
gauge_scheduler_waiting, engine_indexes, model_name
396
)
398
- if envs.VLLM_SERVER_DEV_MODE:
399
- gauge_engine_sleep_state = self._gauge_cls(
400
- name="vllm:engine_sleep_state",
401
- documentation=(
402
- "Engine sleep state; awake = 0 means engine is sleeping; "
403
- "awake = 1 means engine is awake; "
404
- "weights_offloaded = 1 means sleep level 1; "
405
- "discard_all = 1 means sleep level 2."
406
- ),
407
- labelnames=labelnames + ["sleep_state"],
408
- multiprocess_mode="mostrecent",
409
- )
397
411
- self.gauge_engine_sleep_state = {}
412
- sleep_state = ["awake", "weights_offloaded", "discard_all"]
398
+ gauge_engine_sleep_state = self._gauge_cls(
399
+ name="vllm:engine_sleep_state",
400
+ documentation=(
401
+ "Engine sleep state; awake = 0 means engine is sleeping; "
402
+ "awake = 1 means engine is awake; "
403
+ "weights_offloaded = 1 means sleep level 1; "
404
+ "discard_all = 1 means sleep level 2."
405
+ ),
406
+ labelnames=labelnames + ["sleep_state"],
407
+ multiprocess_mode="mostrecent",
408
+ )
409
+
410
+ self.gauge_engine_sleep_state = {}
411
+ sleep_state = ["awake", "weights_offloaded", "discard_all"]
412
414
- for s in sleep_state:
415
- self.gauge_engine_sleep_state[s] = {
416
- idx: gauge_engine_sleep_state.labels(
417
- engine=idx, model_name=model_name, sleep_state=s
418
- )
419
- for idx in engine_indexes
420
- }
413
+ for s in sleep_state:
414
+ self.gauge_engine_sleep_state[s] = {
415
+ idx: gauge_engine_sleep_state.labels(
416
+ engine=idx, model_name=model_name, sleep_state=s
417
+ )
418
+ for idx in engine_indexes
419
+ }
420
422
- # Setting default values
423
- self.record_sleep_state()
421
+ # Setting default values
422
+ self.record_sleep_state()
423
424
# GPU cache
425
#
@@ -1052,9 +1051,6 @@ class PrometheusStatLogger(AggregateStatLoggerBase):
1051
self.gauge_lora_info.labels(**lora_info_labels).set_to_current_time()
1052
1053
def record_sleep_state(self, sleep: int = 0, level: int = 0):
1055
- if not envs.VLLM_SERVER_DEV_MODE:
1056
- return
1057
-
1054
awake = 1
1055
discard_all = 0
1056
weights_offloaded = 0