Merge 525b0ab191 into 044bd824c0

litellm/proxy/auth/auth_object_prefetch.py

@@ -16,7 +16,7 @@ from litellm._internal_context import service_target
 from litellm._logging import verbose_proxy_logger
 from litellm.caching.redis_batch import active_request_redis_batch
 from litellm.caching.redis_cache import RedisCache
-from litellm.constants import DEFAULT_IN_MEMORY_TTL, REGISTRY_ERROR_NEGATIVE_CACHE_TTL
+from litellm.constants import REGISTRY_ERROR_NEGATIVE_CACHE_TTL
 from litellm.models.organization import LiteLLM_OrganizationTable
 from litellm.models.team import LiteLLM_TeamTableCachedObj
 from litellm.models.team_membership import LiteLLM_TeamMembership
@@ -194,14 +194,17 @@ def _iter_entries(refs: AuthObjectRefs, management_ttl: float) -> Iterator[_Cach
             None,
         )
     if refs.organization_id is not None:
+        # Organization entries use the management TTL like every other prefetched object: a
+        # 5s fuse expires before slow requests reach the getters, sending them back to
+        # the DB the prefetch was meant to spare.
         yield _CacheEntry(
-            f"org_id:{refs.organization_id}", "organization_row", LiteLLM_OrganizationTable, DEFAULT_IN_MEMORY_TTL
+            f"org_id:{refs.organization_id}", "organization_row", LiteLLM_OrganizationTable, management_ttl
         )
         yield _CacheEntry(
             f"org_id:{refs.organization_id}:with_budget",
             "organization_row",
             LiteLLM_OrganizationTable,
-            DEFAULT_IN_MEMORY_TTL,
+            management_ttl,
         )
     if refs.project_id is not None:
         yield _CacheEntry(f"project_id:{refs.project_id}", "project_row", LiteLLM_ProjectTableCachedObj, management_ttl)

tests/unit/llms/openai/responses/test_reasoning_effort_capability.py

@@ -0,0 +1,86 @@
+import re
+
+import pytest
+
+import litellm
+from litellm.llms.azure.responses.transformation import AzureOpenAIResponsesAPIConfig
+from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
+
+
+@pytest.mark.parametrize("effort", ["minimal", "low"])
+def test_rejects_explicitly_unsupported_lower_reasoning_effort(effort: str) -> None:
+    config = OpenAIResponsesAPIConfig()
+
+    with pytest.raises(litellm.UnsupportedParamsError, match=f"reasoning.effort={effort}"):
+        config.map_openai_params(
+            response_api_optional_params={"reasoning": {"effort": effort}},
+            model="gpt-5.5-pro",
+            drop_params=False,
+        )
+
+
+def test_keeps_supported_reasoning_effort() -> None:
+    config = OpenAIResponsesAPIConfig()
+
+    result = config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "medium"}},
+        model="gpt-5.5-pro",
+        drop_params=False,
+    )
+
+    assert result["reasoning"] == {"effort": "medium"}
+
+
+def test_drop_params_removes_only_unsupported_effort() -> None:
+    config = OpenAIResponsesAPIConfig()
+
+    result = config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "minimal", "summary": "detailed"}},
+        model="gpt-5.5-pro",
+        drop_params=True,
+    )
+
+    assert result["reasoning"] == {"summary": "detailed"}
+
+
+def test_drop_params_revalidates_temperature_after_effort_drop() -> None:
+    """Dropping an unsupported effort revalidates temperature against the emptied effort.
+
+    `gpt-5.1` really declares `xhigh` unsupported while supporting (and defaulting to)
+    `none`, so no capability stubbing is needed: the dropped effort resolves to `none`
+    and a valid non-default temperature survives.
+    """
+    config = OpenAIResponsesAPIConfig()
+
+    result = config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "xhigh"}, "temperature": 0.5},
+        model="gpt-5.1",
+        drop_params=True,
+    )
+
+    assert "reasoning" not in result
+    assert result["temperature"] == 0.5
+
+
+def test_azure_config_uses_azure_capability_map_for_effort() -> None:
+    """A bare Azure deployment name must resolve against the Azure map entry.
+
+    OpenAI's `gpt-5.4` entry explicitly disables `minimal` while Azure's enables it;
+    the OpenAI lookup would wrongly reject an effort Azure supports.
+    """
+    openai_config = OpenAIResponsesAPIConfig()
+    with pytest.raises(litellm.UnsupportedParamsError, match=re.escape("reasoning.effort=minimal")):
+        openai_config.map_openai_params(
+            response_api_optional_params={"reasoning": {"effort": "minimal"}},
+            model="gpt-5.4",
+            drop_params=False,
+        )
+
+    azure_config = AzureOpenAIResponsesAPIConfig()
+    result = azure_config.map_openai_params(
+        response_api_optional_params={"reasoning": {"effort": "minimal"}},
+        model="gpt-5.4",
+        drop_params=False,
+    )
+
+    assert result["reasoning"] == {"effort": "minimal"}

tests/unit/proxy/auth/test_auth_object_prefetch.py

@@ -186,8 +186,8 @@ async def test_cold_regime_is_one_mget_one_query_and_the_getters_never_touch_io_
     assert sets == sorted(
         [
             f"SET {TEAM_ID}_{USER_ID} ttl=5",
-            f"SET org_id:{ORG_ID} ttl=5",
-            f"SET org_id:{ORG_ID}:with_budget ttl=5",
+            f"SET org_id:{ORG_ID} ttl=60",
+            f"SET org_id:{ORG_ID}:with_budget ttl=60",
             f"SET {USER_ID} ttl=60",
             f"SET team_id:{TEAM_ID} ttl=60",
             f"SET team_membership:{USER_ID}:{TEAM_ID} ttl=None",