Merge 53fdddfc75 into a392a39627

litellm/integrations/opentelemetry.py

@@ -1747,7 +1747,7 @@ class OpenTelemetry(OTELGenAISemconvMixin, CustomLogger):
                 self._token_usage_histogram.record(usage.get("completion_tokens", 0), attributes=out_attrs)
 
         cost: Final = kwargs.get("response_cost")
-        if self._cost_histogram and cost:
+        if self._cost_histogram and cost is not None:
             self._cost_histogram.record(cost, attributes=common_attrs)
 
         # Record latency metrics (TTFT, TPOT, and Total Generation Time)

litellm/integrations/otel/plumbing/metrics.py

@@ -234,7 +234,7 @@ class GenAIMetricRecorder:
             self._record_token_usage(response_obj, common_attrs)
 
         cost: Final = kwargs.get("response_cost")
-        if cost:
+        if cost is not None:
             self._metrics.token_cost.record(cost, attributes=common_attrs)
 
         self._record_time_to_first_token(kwargs, common_attrs)

tests/unit/integrations/otel/test_otel_v2_metrics.py

@@ -193,6 +193,26 @@ def test_all_six_metrics_emitted_when_enabled():
     assert cost_points[0].sum == pytest.approx(RESPONSE_COST)
 
 
+@pytest.mark.parametrize("response_cost", [0, 0.0123, None])
+def test_cost_metric_records_known_values_including_zero(response_cost: float | None):
+    reader = InMemoryMetricReader()
+    logger = _logger(reader, enable_metrics=True)
+    kwargs, response_obj, start, end = _build_call()
+    kwargs["response_cost"] = response_cost
+    asyncio.run(logger.async_log_success_event(kwargs, response_obj, start, end))
+
+    metrics = _metrics_by_name(reader)
+    cost_points = metrics.get(TOKEN_COST, [])
+    if response_cost is None:
+        assert cost_points == []
+        return
+
+    assert len(cost_points) == 1
+    assert cost_points[0].sum == pytest.approx(response_cost)
+    assert cost_points[0].count == 1
+    assert cost_points[0].attributes == metrics[OPERATION_DURATION][0].attributes
+
+
 def test_time_to_first_token_is_streaming_only():
     """time_to_first_token is gated on streaming: a non-streaming call emits the
     other five metrics but never that one."""

tests/unit/integrations/test_opentelemetry.py

@@ -6255,7 +6255,7 @@ class TestOpenTelemetryMetricAttributeFiltering(unittest.TestCase):
             response_obj = json.load(f)
         return kwargs, response_obj
 
-    def _record(self, attributes):
+    def _record(self, attributes, response_cost: float | None = None):
         """Run a real success hook with metrics enabled and return the reader."""
         metric_reader = InMemoryMetricReader()
         meter_provider = MeterProvider(metric_readers=[metric_reader])
@@ -6271,11 +6271,28 @@ class TestOpenTelemetryMetricAttributeFiltering(unittest.TestCase):
         otel.tracer = tracer_provider.get_tracer(__name__)
 
         kwargs, response_obj = self._load_fixtures()
+        if response_cost is not None:
+            kwargs["response_cost"] = response_cost
         start = datetime.utcnow()
         end = start + timedelta(seconds=1)
         otel._handle_success(kwargs, response_obj, start, end)
         return metric_reader
 
+    def test_zero_cost_is_recorded(self):
+        reader = self._record(None, response_cost=0)
+        data = reader.get_metrics_data()
+        cost_metrics = tuple(
+            metric
+            for resource_metrics in data.resource_metrics
+            for scope_metrics in resource_metrics.scope_metrics
+            for metric in scope_metrics.metrics
+            if metric.name == "gen_ai.usage.cost"
+        )
+
+        self.assertEqual(len(cost_metrics), 1)
+        self.assertEqual(cost_metrics[0].data.data_points[0].sum, 0)
+        self.assertEqual(cost_metrics[0].data.data_points[0].count, 1)
+
     def _keysets(self, reader, metric_name):
         """Attribute-key sets, one per recorded data point of `metric_name`."""
         deadline = time.time() + self.POLL_TIMEOUT