runtime: key counter rename rules by the parsed family name

python/kthena/runtime/standard.py

@@ -26,7 +26,13 @@ class EngineType(Enum):
 
 
 class StandardMetricNames:
-    GENERATION_TOKENS_TOTAL = "kthena:generation_tokens_total"
+    # The runtime's parser munges counter families into their OpenMetrics
+    # form: the FAMILY name loses the _total suffix while samples keep it.
+    # Rules are matched by family name, so the counter rule is keyed by the
+    # munged family and RenameMetric's suffix arithmetic restores _total on
+    # the exposed samples, yielding the documented
+    # kthena:generation_tokens_total series.
+    GENERATION_TOKENS = "kthena:generation_tokens"
     NUM_REQUESTS_WAITING = "kthena:num_requests_waiting"
     TIME_TO_FIRST_TOKEN_SECONDS = "kthena:time_to_first_token_seconds"
     TIME_PER_OUTPUT_TOKEN_SECONDS = "kthena:time_per_output_token_seconds"
@@ -36,8 +42,8 @@ class StandardMetricNames:
 STANDARD_RULES: Dict[str, List[MetricOperator]] = {
     EngineType.VLLM.value: [
         RenameMetric(
-            "vllm:generation_tokens_total",
-            StandardMetricNames.GENERATION_TOKENS_TOTAL,
+            "vllm:generation_tokens",
+            StandardMetricNames.GENERATION_TOKENS,
         ),
         RenameMetric(
             "vllm:num_requests_waiting", StandardMetricNames.NUM_REQUESTS_WAITING
@@ -57,8 +63,8 @@ class StandardMetricNames:
     ],
     EngineType.SGLANG.value: [
         RenameMetric(
-            "sglang:generation_tokens_total",
-            StandardMetricNames.GENERATION_TOKENS_TOTAL,
+            "sglang:generation_tokens",
+            StandardMetricNames.GENERATION_TOKENS,
         ),
         RenameMetric("sglang:num_queue_reqs", StandardMetricNames.NUM_REQUESTS_WAITING),
         RenameMetric(

python/kthena/tests/test_metrics.py

@@ -95,3 +95,55 @@ def test_metric_adapter_handles_empty_metrics():
 
     adapter = MetricAdapter(empty_metric_text, standard)
     assert len(adapter.metrics) == 0
+VLLM_COUNTER_METRICS = """
+# HELP vllm:generation_tokens_total Number of generation tokens processed.
+# TYPE vllm:generation_tokens_total counter
+vllm:generation_tokens_total{model_name="m"} 42.0
+""".strip()
+
+SGLANG_COUNTER_METRICS = """
+# HELP sglang:generation_tokens_total Number of generation tokens processed.
+# TYPE sglang:generation_tokens_total counter
+sglang:generation_tokens_total{model_name="m"} 7.0
+""".strip()
+
+
+def _standardized(adapter, name):
+    return [m for m in adapter.metrics if m.name == name]
+
+
+def test_vllm_counter_rename_survives_parser_family_munging():
+    # The parser strips _total from a counter FAMILY name, so the rule must
+    # match the munged family and the samples must come back out with _total.
+    standard = MetricStandard("vllm")
+    adapter = MetricAdapter(VLLM_COUNTER_METRICS, standard)
+
+    twins = _standardized(adapter, "kthena:generation_tokens")
+    assert len(twins) == 1, [m.name for m in adapter.metrics]
+    twin = twins[0]
+    assert twin.type == "counter"
+    sample_names = [s.name for s in twin.samples]
+    assert "kthena:generation_tokens_total" in sample_names
+    total = next(s for s in twin.samples if s.name == "kthena:generation_tokens_total")
+    assert total.value == 42.0
+    assert total.labels == {"model_name": "m"}
+
+
+def test_sglang_counter_rename_survives_parser_family_munging():
+    standard = MetricStandard("sglang")
+    adapter = MetricAdapter(SGLANG_COUNTER_METRICS, standard)
+
+    twins = _standardized(adapter, "kthena:generation_tokens")
+    assert len(twins) == 1, [m.name for m in adapter.metrics]
+    sample_names = [s.name for s in twins[0].samples]
+    assert "kthena:generation_tokens_total" in sample_names
+
+
+def test_process_metrics_exposes_documented_counter_series():
+    import asyncio
+
+    from kthena.runtime.collect import process_metrics
+
+    standard = MetricStandard("vllm")
+    output = asyncio.run(process_metrics(VLLM_COUNTER_METRICS, standard))
+    assert b'kthena:generation_tokens_total{model_name="m"} 42.0' in output