[Metrics] Add `--show-hidden-metrics-for-version` CLI arg (#13295)

docs/source/serving/metrics.md

@@ -36,3 +36,11 @@ The following metrics are exposed:
 :language: python
 :start-after: begin-metrics-definitions
 :::
+
+The following metrics are deprecated and due to be removed in a future version:
+
+- *(No metrics are currently deprecated)*
+
+Note: when metrics are deprecated in version `X.Y`, they are hidden in version `X.Y+1`
+but can be re-enabled using the `--show-hidden-metrics-for-version=X.Y` escape hatch,
+and are then removed in version `X.Y+2`.

tests/test_version.py

+# SPDX-License-Identifier: Apache-2.0
+
+from unittest.mock import patch
+
+import pytest
+
+from vllm import version
+
+
+def test_version_is_defined():
+    assert version.__version__ is not None
+
+
+def test_version_tuple():
+    assert len(version.__version_tuple__) in (3, 4, 5)
+
+
+@pytest.mark.parametrize(
+    "version_tuple, version_str, expected",
+    [
+        ((0, 0, "dev"), "0.0", True),
+        ((0, 0, "dev"), "foobar", True),
+        ((0, 7, 4), "0.6", True),
+        ((0, 7, 4), "0.5", False),
+        ((0, 7, 4), "0.7", False),
+        ((1, 2, 3), "1.1", True),
+        ((1, 2, 3), "1.0", False),
+        ((1, 2, 3), "1.2", False),
+        # This won't work as expected
+        ((1, 0, 0), "1.-1", True),
+        ((1, 0, 0), "0.9", False),
+        ((1, 0, 0), "0.17", False),
+    ])
+def test_prev_minor_version_was(version_tuple, version_str, expected):
+    with patch("vllm.version.__version_tuple__", version_tuple):
+        assert version._prev_minor_version_was(version_str) == expected

vllm/config.py

@@ -2653,7 +2653,9 @@ class DecodingConfig:
 
 @dataclass
 class ObservabilityConfig:
-    """Configuration for observability."""
+    """Configuration for observability - metrics and tracing."""
+    show_hidden_metrics: bool = False
+
     otlp_traces_endpoint: Optional[str] = None
 
     # Collecting detailed timing information for each request can be expensive.

vllm/engine/arg_utils.py

@@ -10,6 +10,7 @@ from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
 import torch
 
 import vllm.envs as envs
+from vllm import version
 from vllm.config import (CacheConfig, CompilationConfig, ConfigFormat,
                          DecodingConfig, DeviceConfig, HfOverrides,
                          KVTransferConfig, LoadConfig, LoadFormat, LoRAConfig,
@@ -188,6 +189,7 @@ class EngineArgs:
     qlora_adapter_name_or_path: Optional[str] = None
     disable_logprobs_during_spec_decoding: Optional[bool] = None
 
+    show_hidden_metrics_for_version: Optional[str] = None
     otlp_traces_endpoint: Optional[str] = None
     collect_detailed_traces: Optional[str] = None
     disable_async_output_proc: bool = False
@@ -909,6 +911,18 @@ class EngineArgs:
                             default=None,
                             help='Name or path of the QLoRA adapter.')
 
+        parser.add_argument('--show-hidden-metrics-for-version',
+                            type=str,
+                            default=None,
+                            help='Enable deprecated Prometheus metrics that '
+                            'have been hidden since the specified version. '
+                            'For example, if a previously deprecated metric '
+                            'has been hidden since the v0.7.0 release, you '
+                            'use --show-hidden-metrics-for-version=0.7 as a '
+                            'temporary escape hatch while you migrate to new '
+                            'metrics. The metric is likely to be removed '
+                            'completely in an upcoming release.')
+
         parser.add_argument(
             '--otlp-traces-endpoint',
             type=str,
@@ -1317,6 +1331,11 @@ class EngineArgs:
         decoding_config = DecodingConfig(
             guided_decoding_backend=self.guided_decoding_backend)
 
+        show_hidden_metrics = False
+        if self.show_hidden_metrics_for_version is not None:
+            show_hidden_metrics = version._prev_minor_version_was(
+                self.show_hidden_metrics_for_version)
+
         detailed_trace_modules = []
         if self.collect_detailed_traces is not None:
             detailed_trace_modules = self.collect_detailed_traces.split(",")
@@ -1326,6 +1345,7 @@ class EngineArgs:
                     f"Invalid module {m} in collect_detailed_traces. "
                     f"Valid modules are {ALLOWED_DETAILED_TRACE_MODULES}")
         observability_config = ObservabilityConfig(
+            show_hidden_metrics=show_hidden_metrics,
             otlp_traces_endpoint=self.otlp_traces_endpoint,
             collect_model_forward_time="model" in detailed_trace_modules
             or "all" in detailed_trace_modules,

vllm/engine/metrics.py

@@ -516,6 +516,11 @@ class PrometheusStatLogger(StatLoggerBase):
         self.metrics = self._metrics_cls(labelnames=list(labels.keys()),
                                          vllm_config=vllm_config)
 
+        # Use this flag to hide metrics that were deprecated in
+        # a previous release and which will be removed future
+        self.show_hidden_metrics = \
+            vllm_config.observability_config.show_hidden_metrics
+
     def _log_gauge(self, gauge, data: Union[int, float]) -> None:
         # Convenience function for logging to gauge.
         gauge.labels(**self.labels).set(data)

vllm/v1/metrics/loggers.py

@@ -95,6 +95,11 @@ class PrometheusStatLogger(StatLoggerBase):
     def __init__(self, vllm_config: VllmConfig):
         self._unregister_vllm_metrics()
 
+        # Use this flag to hide metrics that were deprecated in
+        # a previous release and which will be removed future
+        self.show_hidden_metrics = \
+            vllm_config.observability_config.show_hidden_metrics
+
         labelnames = ["model_name"]
         labelvalues = [vllm_config.model_config.served_model_name]
 

vllm/version.py

@@ -11,3 +11,21 @@ except Exception as e:
 
     __version__ = "dev"
     __version_tuple__ = (0, 0, __version__)
+
+
+def _prev_minor_version_was(version_str):
+    """Check whether a given version matches the previous minor version.
+
+    Return True if version_str matches the previous minor version.
+
+    For example - return True if the current version if 0.7.4 and the
+    supplied version_str is '0.6'.
+
+    Used for --show-hidden-metrics-for-version.
+    """
+    # Match anything if this is a dev tree
+    if __version_tuple__[0:2] == (0, 0):
+        return True
+
+    # Note - this won't do the right thing when we release 1.0!
+    return version_str == f"{__version_tuple__[0]}.{__version_tuple__[1] - 1}"