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 14, 05:30 UTC
[ray][metrics] Replace ':' with '_' for OpenTelemetry compatibility in Ray (#25439)
Signed-off-by: Seiji Eicher <seiji@anyscale.com> Signed-off-by: Seiji Eicher <58963096+eicherseiji@users.noreply.github.com> Co-authored-by: Rui Qiao <161574667+ruisearch42@users.noreply.github.com>
Seiji Eicher committed
Sep 26, 2025 at 09:43 UTC
8d52f2b3a7b75c9efe9eba906ab37780f6e4e5f3
2 files changed
+57
-1
tests/v1/metrics/test_ray_metrics.py
+38
-1
@@ -8,7 +8,8 @@ import ray
8
from vllm.config import ModelDType
9
from vllm.sampling_params import SamplingParams
10
from vllm.v1.engine.async_llm import AsyncEngineArgs, AsyncLLM
11
-from vllm.v1.metrics.ray_wrappers import RayPrometheusStatLogger
11
+from vllm.v1.metrics.ray_wrappers import (RayPrometheusMetric,
12
+ RayPrometheusStatLogger)
13
14
15
@pytest.fixture(scope="function", autouse=True)
@@ -65,3 +66,39 @@ def test_engine_log_metrics_ray(
66
# Create the actor and call the async method
67
actor = EngineTestActor.remote() # type: ignore[attr-defined]
68
ray.get(actor.run.remote())
69
+
70
+
71
+def test_sanitized_opentelemetry_name():
72
+ """Test the metric name sanitization logic for Ray."""
73
+
74
+ # Only a-z, A-Z, 0-9, _, test valid characters are preserved
75
+ valid_name = "valid_metric_123_abcDEF"
76
+ assert RayPrometheusMetric._get_sanitized_opentelemetry_name(
77
+ valid_name) == valid_name
78
+
79
+ # Test dash, dot, are replaced
80
+ name_with_dash_dot = "metric-name.test"
81
+ expected = "metric_name_test"
82
+ assert RayPrometheusMetric._get_sanitized_opentelemetry_name(
83
+ name_with_dash_dot) == expected
84
+
85
+ # Test colon is replaced with underscore
86
+ name_with_colon = "metric:name"
87
+ expected = "metric_name"
88
+ assert RayPrometheusMetric._get_sanitized_opentelemetry_name(
89
+ name_with_colon) == expected
90
+
91
+ # Test multiple invalid characters are replaced
92
+ name_with_invalid = "metric:name@with#special%chars"
93
+ expected = "metric_name_with_special_chars"
94
+ assert RayPrometheusMetric._get_sanitized_opentelemetry_name(
95
+ name_with_invalid) == expected
96
+
97
+ # Test mixed valid and invalid characters
98
+ complex_name = "vllm:engine_stats/time.latency_ms-99p"
99
+ expected = "vllm_engine_stats_time_latency_ms_99p"
100
+ assert RayPrometheusMetric._get_sanitized_opentelemetry_name(
101
+ complex_name) == expected
102
+
103
+ # Test empty string
104
+ assert RayPrometheusMetric._get_sanitized_opentelemetry_name("") == ""
vllm/v1/metrics/ray_wrappers.py
+19
@@ -11,6 +11,7 @@ try:
11
from ray.util.metrics import Metric
12
except ImportError:
13
ray_metrics = None
14
+import regex as re
15
16
17
class RayPrometheusMetric:
@@ -42,6 +43,21 @@ class RayPrometheusMetric:
43
44
return self
45
46
+ @staticmethod
47
+ def _get_sanitized_opentelemetry_name(name: str) -> str:
48
+ """
49
+ For compatibility with Ray + OpenTelemetry, the metric name must be
50
+ sanitized. In particular, this replaces disallowed character (e.g., ':')
51
+ with '_' in the metric name.
52
+ Allowed characters: a-z, A-Z, 0-9, _
53
+
54
+ # ruff: noqa: E501
55
+ Ref: https://github.com/open-telemetry/opentelemetry-cpp/blob/main/sdk/src/metrics/instrument_metadata_validator.cc#L22-L23
56
+ Ref: https://github.com/ray-project/ray/blob/master/src/ray/stats/metric.cc#L107
57
+ """
58
+
59
+ return re.sub(r"[^a-zA-Z0-9_]", "_", name)
60
+
61
62
class RayGaugeWrapper(RayPrometheusMetric):
63
"""Wraps around ray.util.metrics.Gauge to provide same API as
@@ -58,6 +74,7 @@ class RayGaugeWrapper(RayPrometheusMetric):
74
# implemented at the observability layer (Prometheus/Grafana).
75
del multiprocess_mode
76
labelnames_tuple = tuple(labelnames) if labelnames else None
77
+ name = self._get_sanitized_opentelemetry_name(name)
78
self.metric = ray_metrics.Gauge(name=name,
79
description=documentation,
80
tag_keys=labelnames_tuple)
@@ -79,6 +96,7 @@ class RayCounterWrapper(RayPrometheusMetric):
96
documentation: Optional[str] = "",
97
labelnames: Optional[list[str]] = None):
98
labelnames_tuple = tuple(labelnames) if labelnames else None
99
+ name = self._get_sanitized_opentelemetry_name(name)
100
self.metric = ray_metrics.Counter(name=name,
101
description=documentation,
102
tag_keys=labelnames_tuple)
@@ -99,6 +117,7 @@ class RayHistogramWrapper(RayPrometheusMetric):
117
labelnames: Optional[list[str]] = None,
118
buckets: Optional[list[float]] = None):
119
labelnames_tuple = tuple(labelnames) if labelnames else None
120
+ name = self._get_sanitized_opentelemetry_name(name)
121
boundaries = buckets if buckets else []
122
self.metric = ray_metrics.Histogram(name=name,
123
description=documentation,