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 15, 22:15 UTC
[Metrics] Add `--show-hidden-metrics-for-version` CLI arg (#13295)
Mark McLoughlin committed
Feb 22, 2025 at 08:20 UTC
2cb8c1540e27ffebdb668a8f10ec7b8b7703aab3
7 files changed
+95
-1
docs/source/serving/metrics.md
+8
@@ -36,3 +36,11 @@ The following metrics are exposed:
36
:language: python
37
:start-after: begin-metrics-definitions
38
:::
39
+
40
+The following metrics are deprecated and due to be removed in a future version:
41
+
42
+- *(No metrics are currently deprecated)*
43
+
44
+Note: when metrics are deprecated in version `X.Y`, they are hidden in version `X.Y+1`
45
+but can be re-enabled using the `--show-hidden-metrics-for-version=X.Y` escape hatch,
46
+and are then removed in version `X.Y+2`.
tests/test_version.py
new
+36
@@ -0,0 +1,36 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+
3
+from unittest.mock import patch
4
+
5
+import pytest
6
+
7
+from vllm import version
8
+
9
+
10
+def test_version_is_defined():
11
+ assert version.__version__ is not None
12
+
13
+
14
+def test_version_tuple():
15
+ assert len(version.__version_tuple__) in (3, 4, 5)
16
+
17
+
18
+@pytest.mark.parametrize(
19
+ "version_tuple, version_str, expected",
20
+ [
21
+ ((0, 0, "dev"), "0.0", True),
22
+ ((0, 0, "dev"), "foobar", True),
23
+ ((0, 7, 4), "0.6", True),
24
+ ((0, 7, 4), "0.5", False),
25
+ ((0, 7, 4), "0.7", False),
26
+ ((1, 2, 3), "1.1", True),
27
+ ((1, 2, 3), "1.0", False),
28
+ ((1, 2, 3), "1.2", False),
29
+ # This won't work as expected
30
+ ((1, 0, 0), "1.-1", True),
31
+ ((1, 0, 0), "0.9", False),
32
+ ((1, 0, 0), "0.17", False),
33
+ ])
34
+def test_prev_minor_version_was(version_tuple, version_str, expected):
35
+ with patch("vllm.version.__version_tuple__", version_tuple):
36
+ assert version._prev_minor_version_was(version_str) == expected
vllm/config.py
+3
-1
@@ -2653,7 +2653,9 @@ class DecodingConfig:
2653
2654
@dataclass
2655
class ObservabilityConfig:
2656
- """Configuration for observability."""
2656
+ """Configuration for observability - metrics and tracing."""
2657
+ show_hidden_metrics: bool = False
2658
+
2659
otlp_traces_endpoint: Optional[str] = None
2660
2661
# Collecting detailed timing information for each request can be expensive.
vllm/engine/arg_utils.py
+20
@@ -10,6 +10,7 @@ from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
10
import torch
11
12
import vllm.envs as envs
13
+from vllm import version
14
from vllm.config import (CacheConfig, CompilationConfig, ConfigFormat,
15
DecodingConfig, DeviceConfig, HfOverrides,
16
KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
@@ -188,6 +189,7 @@ class EngineArgs:
189
qlora_adapter_name_or_path: Optional[str] = None
190
disable_logprobs_during_spec_decoding: Optional[bool] = None
191
192
+ show_hidden_metrics_for_version: Optional[str] = None
193
otlp_traces_endpoint: Optional[str] = None
194
collect_detailed_traces: Optional[str] = None
195
disable_async_output_proc: bool = False
@@ -909,6 +911,18 @@ class EngineArgs:
911
default=None,
912
help='Name or path of the QLoRA adapter.')
913
914
+ parser.add_argument('--show-hidden-metrics-for-version',
915
+ type=str,
916
+ default=None,
917
+ help='Enable deprecated Prometheus metrics that '
918
+ 'have been hidden since the specified version. '
919
+ 'For example, if a previously deprecated metric '
920
+ 'has been hidden since the v0.7.0 release, you '
921
+ 'use --show-hidden-metrics-for-version=0.7 as a '
922
+ 'temporary escape hatch while you migrate to new '
923
+ 'metrics. The metric is likely to be removed '
924
+ 'completely in an upcoming release.')
925
+
926
parser.add_argument(
927
'--otlp-traces-endpoint',
928
type=str,
@@ -1317,6 +1331,11 @@ class EngineArgs:
1331
decoding_config = DecodingConfig(
1332
guided_decoding_backend=self.guided_decoding_backend)
1333
1334
+ show_hidden_metrics = False
1335
+ if self.show_hidden_metrics_for_version is not None:
1336
+ show_hidden_metrics = version._prev_minor_version_was(
1337
+ self.show_hidden_metrics_for_version)
1338
+
1339
detailed_trace_modules = []
1340
if self.collect_detailed_traces is not None:
1341
detailed_trace_modules = self.collect_detailed_traces.split(",")
@@ -1326,6 +1345,7 @@ class EngineArgs:
1345
f"Invalid module {m} in collect_detailed_traces. "
1346
f"Valid modules are {ALLOWED_DETAILED_TRACE_MODULES}")
1347
observability_config = ObservabilityConfig(
1348
+ show_hidden_metrics=show_hidden_metrics,
1349
otlp_traces_endpoint=self.otlp_traces_endpoint,
1350
collect_model_forward_time="model" in detailed_trace_modules
1351
or "all" in detailed_trace_modules,
vllm/engine/metrics.py
+5
@@ -516,6 +516,11 @@ class PrometheusStatLogger(StatLoggerBase):
516
self.metrics = self._metrics_cls(labelnames=list(labels.keys()),
517
vllm_config=vllm_config)
518
519
+ # Use this flag to hide metrics that were deprecated in
520
+ # a previous release and which will be removed future
521
+ self.show_hidden_metrics = \
522
+ vllm_config.observability_config.show_hidden_metrics
523
+
524
def _log_gauge(self, gauge, data: Union[int, float]) -> None:
525
# Convenience function for logging to gauge.
526
gauge.labels(**self.labels).set(data)
vllm/v1/metrics/loggers.py
+5
@@ -95,6 +95,11 @@ class PrometheusStatLogger(StatLoggerBase):
95
def __init__(self, vllm_config: VllmConfig):
96
self._unregister_vllm_metrics()
97
98
+ # Use this flag to hide metrics that were deprecated in
99
+ # a previous release and which will be removed future
100
+ self.show_hidden_metrics = \
101
+ vllm_config.observability_config.show_hidden_metrics
102
+
103
labelnames = ["model_name"]
104
labelvalues = [vllm_config.model_config.served_model_name]
105
vllm/version.py
+18
@@ -11,3 +11,21 @@ except Exception as e:
11
12
__version__ = "dev"
13
__version_tuple__ = (0, 0, __version__)
14
+
15
+
16
+def _prev_minor_version_was(version_str):
17
+ """Check whether a given version matches the previous minor version.
18
+
19
+ Return True if version_str matches the previous minor version.
20
+
21
+ For example - return True if the current version if 0.7.4 and the
22
+ supplied version_str is '0.6'.
23
+
24
+ Used for --show-hidden-metrics-for-version.
25
+ """
26
+ # Match anything if this is a dev tree
27
+ if __version_tuple__[0:2] == (0, 0):
28
+ return True
29
+
30
+ # Note - this won't do the right thing when we release 1.0!
31
+ return version_str == f"{__version_tuple__[0]}.{__version_tuple__[1] - 1}"