Remove unregister_vllm_metrics no-op patch (#4396)

Signed-off-by: vraiti <vraiti@redhat.com>
Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>

tests/metrics/test_prometheus.py

@@ -12,10 +12,10 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 _MODEL = "test-model"
 
 _PIPELINE_METRICS = [
-    "vllm:omni_num_requests_running",
-    "vllm:omni_num_requests_waiting",
-    "vllm:omni_requests_success",
-    "vllm:omni_e2e_request_latency_s",
+    "vllm_omni:num_requests_running",
+    "vllm_omni:num_requests_waiting",
+    "vllm_omni:requests_success",
+    "vllm_omni:e2e_request_latency_s",
 ]
 
 
@@ -58,32 +58,32 @@ class TestMetricObservation:
         # Per-reason buckets sourced from the merged completion Counter.
         stop = _sample_value(
             scrape_output,
-            f'vllm:omni_requests_success_total{{finished_reason="stop",model_name="{_MODEL}"}}',
+            f'vllm_omni:requests_success_total{{finished_reason="stop",model_name="{_MODEL}"}}',
         )
         assert stop == 2.0
 
         length = _sample_value(
             scrape_output,
-            f'vllm:omni_requests_success_total{{finished_reason="length",model_name="{_MODEL}"}}',
+            f'vllm_omni:requests_success_total{{finished_reason="length",model_name="{_MODEL}"}}',
         )
         assert length == 1.0
 
         abort = _sample_value(
             scrape_output,
-            f'vllm:omni_requests_success_total{{finished_reason="abort",model_name="{_MODEL}"}}',
+            f'vllm_omni:requests_success_total{{finished_reason="abort",model_name="{_MODEL}"}}',
         )
         assert abort == 1.0
 
     def test_gauge_values(self, scrape_output: str) -> None:
         running = _sample_value(
             scrape_output,
-            f'vllm:omni_num_requests_running{{model_name="{_MODEL}"}}',
+            f'vllm_omni:num_requests_running{{model_name="{_MODEL}"}}',
         )
         assert running == 5.0
 
         waiting = _sample_value(
             scrape_output,
-            f'vllm:omni_num_requests_waiting{{model_name="{_MODEL}"}}',
+            f'vllm_omni:num_requests_waiting{{model_name="{_MODEL}"}}',
         )
         assert waiting == 2.0
 
@@ -93,7 +93,7 @@ class TestMetricObservation:
         # observing the latency histogram, so the count stays at 3.
         e2e_count = _sample_value(
             scrape_output,
-            f'vllm:omni_e2e_request_latency_s_count{{model_name="{_MODEL}"}}',
+            f'vllm_omni:e2e_request_latency_s_count{{model_name="{_MODEL}"}}',
         )
         assert e2e_count == 3.0
 
@@ -110,8 +110,8 @@ class TestLabelCorrectness:
     def test_no_legacy_seconds_or_ms_families(self, scrape_output: str) -> None:
         # Renamed: *_time_ms → *_s; *_time_seconds dropped.
         for legacy in (
-            "vllm:omni_request_queue_time_seconds",
-            "vllm:omni_e2e_request_latency_seconds",
+            "vllm_omni:request_queue_time_seconds",
+            "vllm_omni:e2e_request_latency_seconds",
         ):
             assert legacy not in scrape_output, f"legacy family {legacy} still registered"
 
@@ -170,8 +170,8 @@ class TestRequestLifecycleGauges:
         assert obj.engine._running_counter.value == 0
         assert len(obj.request_states) == 0
         out = generate_latest(registry).decode()
-        assert _sample_value(out, 'vllm:omni_num_requests_running{model_name="lifecycle-test"}') == 0.0
-        assert _sample_value(out, 'vllm:omni_num_requests_waiting{model_name="lifecycle-test"}') == 0.0
+        assert _sample_value(out, 'vllm_omni:num_requests_running{model_name="lifecycle-test"}') == 0.0
+        assert _sample_value(out, 'vllm_omni:num_requests_waiting{model_name="lifecycle-test"}') == 0.0
 
     def test_gauges_reflect_remaining_requests_after_one_completes(self, registry: CollectorRegistry) -> None:
         from types import SimpleNamespace
@@ -196,5 +196,5 @@ class TestRequestLifecycleGauges:
         obj._log_summary_and_cleanup("req-1")
 
         out = generate_latest(registry).decode()
-        assert _sample_value(out, 'vllm:omni_num_requests_running{model_name="lifecycle-test-2"}') == 1.0
-        assert _sample_value(out, 'vllm:omni_num_requests_waiting{model_name="lifecycle-test-2"}') == 0.0
+        assert _sample_value(out, 'vllm_omni:num_requests_running{model_name="lifecycle-test-2"}') == 1.0
+        assert _sample_value(out, 'vllm_omni:num_requests_waiting{model_name="lifecycle-test-2"}') == 0.0

vllm_omni/metrics/definitions.py

@@ -13,9 +13,9 @@ time-bearing metrics use the ``_s`` suffix (values in seconds), counters use
 
 from typing import Any
 
-# vllm:omni_ avoids upstream's unregister_vllm_metrics() stripping, which
-# removes every collector whose ``_name`` does not start with ``vllm``.
-METRIC_PREFIX = "vllm:omni_"
+# vllm_omni: namespace for omni-specific Prometheus families, distinct from
+# the upstream vllm:* families.
+METRIC_PREFIX = "vllm_omni:"
 
 
 # ============================================================================

vllm_omni/metrics/prometheus.py

@@ -42,9 +42,8 @@ _generation_tokens_family = Counter(
 class OmniPrometheusMetrics:
     """Label-bound wrapper around the raw Prometheus metrics.
 
-    Metric collectors use the ``vllm:omni_`` prefix to avoid being
-    removed by upstream vLLM's ``unregister_vllm_metrics()``, which
-    strips every collector whose ``_name`` contains ``"vllm"``.
+    Metric collectors use the ``vllm_omni:`` prefix, distinct from the
+    upstream ``vllm:*`` families.
     """
 
     def __init__(self, model_name: str, log_stats: bool = True) -> None:

vllm_omni/patch.py

@@ -313,43 +313,6 @@ for module_name, module in list(sys.modules.items()):
     if hasattr(module, "EngineCoreRequest") and module.EngineCoreRequest == _OriginalEngineCoreRequest:
         module.EngineCoreRequest = OmniEngineCoreRequest
 
-# =============================================================================
-# Patch unregister_vllm_metrics to skip vllm:omni_* collectors
-# =============================================================================
-# WHY: Upstream unregister_vllm_metrics() uses `"vllm" in collector._name` to
-# strip collectors before each new PrometheusStatLogger registers, which is
-# how it avoids "Duplicated timeseries" when the same process spawns multiple
-# engines / orchestrators.  But its substring match also wipes our vllm:omni_*
-# families, so we must replace it with a scoped version that keeps ours.
-#
-# REMOVAL: Remove this patch once upstream vLLM adds
-# _STAT_LOGGER_METRIC_NAMES to vllm.v1.metrics.prometheus and scopes
-# unregister_vllm_metrics() to that set.  Track:
-# https://github.com/vllm-project/vllm/pull/42331
-import vllm.v1.metrics.prometheus as _vllm_prometheus  # noqa: E402
-
-_logger = logging.getLogger(__name__)
-
-
-def _scoped_unregister_vllm_metrics():
-    """Drop upstream vllm:* collectors but preserve vllm:omni_* (ours)."""
-    from prometheus_client import REGISTRY
-
-    for collector in list(REGISTRY._collector_to_names):
-        name = getattr(collector, "_name", "")
-        if "vllm" not in name:
-            continue
-        if name.startswith("vllm:omni_") or name.startswith("vllm_omni"):
-            continue
-        REGISTRY.unregister(collector)
-
-
-_vllm_prometheus.unregister_vllm_metrics = _scoped_unregister_vllm_metrics
-_logger.warning(
-    "Monkey-patched unregister_vllm_metrics() to scope drops to non-omni vllm:* collectors. "
-    "Remove this patch once vLLM adds _STAT_LOGGER_METRIC_NAMES."
-)
-
 
 # Patch: add qwen3_omni_moe to vllm's chat template fallback registry.
 # Qwen/Qwen3-Omni-30B-A3B-Instruct stores its chat_template in a standalone