Merge 0041a3708f into 527d882508

litellm/integrations/otel/__init__.py

@@ -58,6 +58,7 @@ from litellm.integrations.otel.model.semconv import (
     Metric,
     Network,
     NetworkTransport,
+    OpenAI,
     RpcSystem,
     Server,
     resolve_operation,
@@ -102,6 +103,7 @@ __all__ = [
     "Metric",
     "Network",
     "NetworkTransport",
+    "OpenAI",
     "OpenTelemetryV2Config",
     "ProxyRequestSpanData",
     "RequestContext",

litellm/integrations/otel/mappers/genai.py

@@ -34,6 +34,7 @@ from litellm.integrations.otel.model.semconv import (
     GenAI,
     JsonRpc,
     LiteLLM,
+    OpenAI,
     RpcSystem,
     Server,
 )
@@ -63,6 +64,8 @@ class GenAIMapper:
         GenAI.RESPONSE_ID: lambda d: d.response_id,
         GenAI.RESPONSE_FINISH_REASONS: lambda d: list(d.finish_reasons) if d.finish_reasons else None,
         GenAI.RESPONSE_TIME_TO_FIRST_CHUNK: lambda d: d.time_to_first_chunk_seconds,
+        OpenAI.REQUEST_SERVICE_TIER: lambda d: d.requested_service_tier,
+        OpenAI.RESPONSE_SERVICE_TIER: lambda d: d.served_service_tier,
         GenAI.USAGE_INPUT_TOKENS: lambda d: d.usage.input_tokens,
         GenAI.USAGE_OUTPUT_TOKENS: lambda d: d.usage.output_tokens,
         GenAI.USAGE_CACHE_CREATION_INPUT_TOKENS: lambda d: d.usage.cache_creation_input_tokens,

litellm/integrations/otel/model/payloads.py

@@ -36,6 +36,7 @@ from litellm.integrations.otel.model.utils import (
     as_str_tuple,
 )
 from litellm.integrations.otel.routing import RoutingAttributeValue, routing_decision_attributes
+from litellm.litellm_core_utils.service_tier_utils import get_requested_service_tier, get_served_service_tier
 
 # ``RequestIdentity`` and the request-metadata translation now live in
 # :mod:`metadata`; re-exported here so existing ``model.payloads`` imports keep
@@ -450,6 +451,8 @@ class LLMCallSpanData:
     routing_attributes: Mapping[str, RoutingAttributeValue] = field(default_factory=lambda: MappingProxyType({}))
     cache_hit: bool | None = None
     saved_cache_cost: float | None = None
+    requested_service_tier: str | None = None
+    served_service_tier: str | None = None
 
     @classmethod
     def from_standard_logging_payload(
@@ -531,6 +534,8 @@ class LLMCallSpanData:
             routing_attributes=routing_decision_attributes(decision),
             cache_hit=cache_hit,
             saved_cache_cost=as_float(payload.get("saved_cache_cost")) if cache_hit else None,  # pyright: ignore[reportUnknownMemberType]  # TypedDict .get on a str key is partially unknown
+            requested_service_tier=get_requested_service_tier(payload),
+            served_service_tier=get_served_service_tier(payload),
         )
 
 

litellm/integrations/otel/model/semconv.py

@@ -133,6 +133,14 @@ class GenAI:
     PROMPT_NAME: Final = "gen_ai.prompt.name"
 
 
+class OpenAI:
+    """OpenAI-namespace keys from the GenAI semconv registry. Carries the service tier for every
+    provider, not only OpenAI."""
+
+    REQUEST_SERVICE_TIER: Final = "openai.request.service_tier"
+    RESPONSE_SERVICE_TIER: Final = "openai.response.service_tier"
+
+
 class MCP:
     """OTel GenAI MCP (Model Context Protocol) span-attribute keys.
 

tests/unit/integrations/otel/test_otel_v2_emitter.py

@@ -183,6 +183,64 @@ def test_llm_call_span_golden():
     assert span.status.status_code is StatusCode.UNSET
 
 
+_REQUEST_TIER: Final = "openai.request.service_tier"
+_RESPONSE_TIER: Final = "openai.response.service_tier"
+
+
+@pytest.mark.parametrize(
+    ("overrides", "expected_request_tier", "expected_response_tier"),
+    [
+        pytest.param(
+            {
+                "model_parameters": {"service_tier": "flex"},
+                "response": {"id": "resp_1", "service_tier": "flex", "choices": [{"finish_reason": "stop"}]},
+            },
+            "flex",
+            "flex",
+            id="requested-and-served",
+        ),
+        pytest.param(
+            {
+                "model_parameters": {"service_tier": "auto"},
+                "response": {"id": "resp_1", "service_tier": "priority", "choices": [{"finish_reason": "stop"}]},
+            },
+            "auto",
+            "priority",
+            id="provider-picks-the-tier",
+        ),
+        pytest.param(
+            {"metadata": {"usage_object": {"service_tier": "standard"}}},
+            None,
+            "standard",
+            id="served-tier-on-usage",
+        ),
+        pytest.param(
+            {"model_parameters": {"service_tier": "priority"}},
+            "priority",
+            None,
+            id="requested-only",
+        ),
+        pytest.param(
+            {"model_parameters": {"service_tier": "not-a-tier"}},
+            None,
+            None,
+            id="unknown-requested-tier",
+        ),
+        pytest.param({}, None, None, id="no-tier"),
+    ],
+)
+def test_llm_call_span_service_tier(
+    overrides: Mapping[str, object], expected_request_tier: str | None, expected_response_tier: str | None
+) -> None:
+    engine, exporter = _engine()
+
+    engine.emit(SpanRole.LLM_CALL, LLMCallSpanData.from_standard_logging_payload(_payload(**overrides)))
+
+    (span,) = exporter.get_finished_spans()
+    assert span.attributes.get(_REQUEST_TIER) == expected_request_tier
+    assert span.attributes.get(_RESPONSE_TIER) == expected_response_tier
+
+
 def test_legacy_dual_emit_on():
     engine, exporter = _engine(legacy_compat=True)
     engine.emit(SpanRole.LLM_CALL, LLMCallSpanData.from_standard_logging_payload(_payload()))

tests/unit/integrations/otel/test_otel_v2_sources_of_truth.py

@@ -159,11 +159,11 @@ def _all_constants(cls):
 
 
 def test_attribute_keys_are_unique_across_namespaces():
-    from litellm.integrations.otel import MCP, Client, JsonRpc, LiteLLMError, Network
+    from litellm.integrations.otel import MCP, Client, JsonRpc, LiteLLMError, Network, OpenAI
 
     # prefixes are allowed to be substrings; exact keys must not collide.
     exact = set()
-    for cls in (GenAI, Error, LiteLLMError, Server, HTTP, DB, MCP, JsonRpc, Network, Client):
+    for cls in (GenAI, Error, LiteLLMError, Server, HTTP, DB, MCP, JsonRpc, Network, Client, OpenAI):
         for key in _all_constants(cls):
             assert key not in exact, f"duplicate attribute key {key}"
             exact.add(key)