fix(otel): promote litellm.request.model via baggage instead of gen_ai.request.model

litellm/integrations/otel/model/baggage.py

@@ -32,7 +32,7 @@
     ),
     LiteLLM.KEY_HASH: lambda identity, model, team_metadata_keys: identity.key_hash,
     LiteLLM.END_USER: lambda identity, model, team_metadata_keys: identity.end_user,
-    GenAI.REQUEST_MODEL: lambda identity, model, team_metadata_keys: model,
+    LiteLLM.REQUEST_MODEL: lambda identity, model, team_metadata_keys: model,
     LiteLLM.PROVIDER_MODEL: lambda identity, model, team_metadata_keys: identity.provider_model,
 }
 
@@ -44,7 +44,7 @@
     LiteLLM.TEAM_ALIAS,
     LiteLLM.TEAM_METADATA,
     LiteLLM.KEY_HASH,
-    GenAI.REQUEST_MODEL,
+    LiteLLM.REQUEST_MODEL,
     LiteLLM.PROVIDER_MODEL,
 )
 
@@ -65,6 +65,17 @@
 DEFAULT_BAGGAGE_TEAM_METADATA_KEYS: Final[tuple[str, ...]] = ()
 
 
+def _normalize_promoted_keys(promoted_keys: tuple[str, ...]) -> tuple[str, ...]:
+    """Remap legacy ``gen_ai.request.model`` allowlist entries to the vendor key.
+
+    Canonical ``gen_ai.request.model`` must not ride Baggage onto non-LLM spans.
+    Deployments that still list it in ``baggage_promoted_keys`` keep model
+    correlation via ``litellm.request.model`` instead of silently dropping it.
+    """
+    remapped = tuple(LiteLLM.REQUEST_MODEL if key == GenAI.REQUEST_MODEL else key for key in promoted_keys)
+    return tuple(key for i, key in enumerate(remapped) if key not in remapped[:i])
+
+
 def promoted_baggage(
     identity: RequestIdentity,
     request_model: str | None,
@@ -79,6 +90,7 @@ def promoted_baggage(
     ``team_metadata_keys`` selects sub-keys of the team's metadata to promote
     under ``litellm.team.metadata``. Empty values are dropped.
     """
+    promoted_keys = _normalize_promoted_keys(promoted_keys)
     out: dict[str, str] = {}
     for key, extract in _PROMOTABLE.items():
         if key in promoted_keys:

litellm/integrations/otel/model/semconv.py

@@ -229,8 +229,12 @@ class LiteLLM:
     TEAM_METADATA: Final = "litellm.team.metadata"
     KEY_HASH: Final = "litellm.api_key.hash"
     END_USER: Final = "litellm.end_user.id"
+    # User-facing model from the inbound request. Promoted via Baggage for
+    # correlation on non-GenAI spans; ``gen_ai.request.model`` is stamped only
+    # on LLM-call spans by the GenAI mapper.
+    REQUEST_MODEL: Final = "litellm.request.model"
     # The model string litellm actually sent to the provider (the deployment's
-    # ``litellm_params.model``), distinct from the user-facing ``gen_ai.request.model``.
+    # ``litellm_params.model``), distinct from the user-facing request model.
     PROVIDER_MODEL: Final = "litellm.provider.model"
     REQUEST_STREAMING: Final = "litellm.request.streaming"
     GUARDRAIL_NAME: Final = "litellm.guardrail.name"

tests/test_litellm/integrations/otel/test_otel_v2_baggage.py

@@ -74,7 +74,33 @@ def test_identity_promoted_onto_every_span():
     for span in spans:
         assert span.attributes.get(LiteLLM.TEAM_ID) == "t1"
         assert span.attributes.get(LiteLLM.TEAM_ALIAS) == "team one"
-        assert span.attributes.get(GenAI.REQUEST_MODEL) == "gpt-4o"
+        assert span.attributes.get(LiteLLM.REQUEST_MODEL) == "gpt-4o"
+    for span in spans:
+        if span.name == "chat gpt-4o":
+            assert span.attributes.get(GenAI.REQUEST_MODEL) == "gpt-4o"
+        else:
+            assert GenAI.REQUEST_MODEL not in span.attributes
+
+
+def test_gen_ai_request_model_not_promoted_via_baggage():
+    """Baggage must not stamp canonical gen_ai.request.model on non-LLM spans."""
+    engine, exporter = _engine_and_exporter()
+    data = LLMCallSpanData.from_standard_logging_payload(_payload())
+    bag = promoted_baggage(data.identity, data.request_model, BAGGAGE_PROMOTED_KEYS)
+    ctx = ctx_mod.set_request_baggage(bag)
+
+    root = engine.start_span(SpanRole.PROXY_REQUEST, "POST /chat/completions", ctx)
+    root_ctx = ctx_mod.context_from_span(root, ctx)
+    engine.emit(SpanRole.SERVICE, ServiceSpanData("redis", call_type="set"), root_ctx)
+    root.end()
+
+    spans = exporter.get_finished_spans()
+    service_span = next(s for s in spans if s.name != "POST /chat/completions")
+    assert service_span.attributes.get(LiteLLM.REQUEST_MODEL) == "gpt-4o"
+    assert GenAI.REQUEST_MODEL not in service_span.attributes
+    root_span = next(s for s in spans if s.name == "POST /chat/completions")
+    assert root_span.attributes.get(LiteLLM.REQUEST_MODEL) == "gpt-4o"
+    assert GenAI.REQUEST_MODEL not in root_span.attributes
 
 
 def test_team_metadata_promoted_only_for_allowlisted_subkeys():
@@ -99,7 +125,23 @@ def test_team_metadata_promoted_only_for_allowlisted_subkeys():
     assert json.loads(span.attributes[LiteLLM.TEAM_METADATA]) == {"tier": "gold"}
     # provider model is distinct from the user-facing request model
     assert span.attributes.get(LiteLLM.PROVIDER_MODEL) == "azure/my-deployment"
-    assert span.attributes.get(GenAI.REQUEST_MODEL) == "gpt-4o"
+    assert span.attributes.get(LiteLLM.REQUEST_MODEL) == "gpt-4o"
+    assert GenAI.REQUEST_MODEL not in span.attributes
+
+
+def test_legacy_gen_ai_request_model_allowlist_remaps_to_vendor_key():
+    """Explicit baggage_promoted_keys with gen_ai.request.model still promote
+    the model, but under litellm.request.model so non-LLM spans are not
+    misclassified as GenAI operations."""
+    data = LLMCallSpanData.from_standard_logging_payload(_payload())
+    bag = promoted_baggage(
+        data.identity,
+        data.request_model,
+        (LiteLLM.TEAM_ID, GenAI.REQUEST_MODEL),
+    )
+    assert bag.get(LiteLLM.REQUEST_MODEL) == "gpt-4o"
+    assert GenAI.REQUEST_MODEL not in bag
+    assert bag.get(LiteLLM.TEAM_ID) == "t1"
 
 
 def test_team_metadata_not_promoted_by_default():

tests/test_litellm/integrations/otel/test_otel_v2_sources_of_truth.py

@@ -698,7 +698,7 @@ def test_promoted_baggage_is_bounded_allowlist():
     promoted = promoted_baggage(identity, "gpt-4o", BAGGAGE_PROMOTED_KEYS)
     assert promoted[LiteLLM.TEAM_ID] == "t1"
     assert promoted[LiteLLM.TEAM_ALIAS] == "team one"
-    assert promoted[GenAI.REQUEST_MODEL] == "gpt-4o"
+    assert promoted[LiteLLM.REQUEST_MODEL] == "gpt-4o"
     # allowlisted metadata sub-key is promoted under the litellm.metadata.* prefix
     assert promoted[f"{LiteLLM.METADATA_PREFIX}user_api_key_org_id"] == "org1"
     # full metadata blob is NOT promoted