Merge 75f7ccc8b1 into 3f38730d2c

litellm/llms/oci/chat/transformation.py

@@ -73,7 +73,6 @@ from litellm.types.utils import (
     ModelResponse,
     ModelResponseStream,
 )
-from litellm.utils import supports_reasoning
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
@@ -96,16 +95,15 @@ def _model_uses_max_completion_tokens(model: str) -> bool:
     and accept ``maxCompletionTokens`` everywhere, so route the whole vendor
     prefix to it rather than chasing each new release in
     ``model_prices_and_context_window.json``. The ``openai.gpt-oss-*`` open
-    weights are served by OCI's own stack and keep ``maxTokens``. Any other
-    vendor falls back to the catalog's ``supports_reasoning`` flag.
+    weights are served by OCI's own stack and keep ``maxTokens``, as does every
+    other vendor whatever its catalog reasoning flag: OCI's Gemini models ignore
+    ``maxCompletionTokens`` and fall back to a default of about 4K output tokens.
     """
     if not model:
         return False
     name: Final = model[4:] if model.lower().startswith("oci/") else model
     lowered: Final = name.lower()
-    if lowered.startswith("openai."):
-        return not lowered.startswith("openai.gpt-oss")
-    return supports_reasoning(model=name, custom_llm_provider="oci")
+    return lowered.startswith("openai.") and not lowered.startswith("openai.gpt-oss")
 
 
 def _iter_sse_events(stream: Iterator[str]) -> Iterator[str]:
@@ -572,8 +570,7 @@ class OCIChatConfig(BaseConfig):
 
         # OpenAI reasoning models on OCI (e.g. GPT-5 family) reject "maxTokens"
         # and require "maxCompletionTokens" per OCI's /20231130/Chat schema.
-        # Driven by the supports_reasoning flag in the model catalog. Cohere's
-        # endpoint uses "maxTokens" regardless, so the override is GENERIC-only.
+        # Cohere's endpoint uses "maxTokens" regardless, so the override is GENERIC-only.
         max_tokens_key: Final = (
             "maxCompletionTokens"
             if vendor != OCIVendors.COHERE and model and _model_uses_max_completion_tokens(model)

tests/unit/llms/oci/chat/test_oci_generic_chat.py

@@ -8,6 +8,7 @@ from unittest.mock import MagicMock
 import httpx
 
 from litellm import ModelResponse
+from litellm.constants import DEFAULT_OCI_CHAT_MAX_TOKENS
 from litellm.llms.oci.chat.generic import (
     adapt_messages_to_generic_oci_standard,
     adapt_messages_to_generic_oci_standard_content_message,
@@ -369,6 +370,27 @@ def _register_oci_gpt5_in_catalog():
         litellm.model_cost.pop(key, None)
 
 
+@pytest.fixture
+def _flag_oci_gemini_as_reasoning_in_catalog():
+    """Mark OCI's Gemini models as reasoning models, as the bundled catalog has done since #41902,
+    whichever map ``litellm.model_cost`` was populated from."""
+    import litellm
+
+    keys = ("oci/google.gemini-2.5-pro", "oci/google.gemini-2.5-flash")
+    originals = {key: litellm.model_cost.get(key) for key in keys}
+    for key in keys:
+        litellm.model_cost[key] = {
+            **(originals[key] or {"litellm_provider": "oci", "mode": "chat"}),
+            "supports_reasoning": True,
+        }
+    yield
+    for key, original in originals.items():
+        if original is None:
+            litellm.model_cost.pop(key, None)
+        else:
+            litellm.model_cost[key] = original
+
+
 class TestGpt5MaxCompletionTokens:
     def test_helper_detects_gpt5_family(self, _register_oci_gpt5_in_catalog):
         assert _model_uses_max_completion_tokens("openai.gpt-5") is True
@@ -441,6 +463,18 @@ class TestGpt5MaxCompletionTokens:
         assert out.get("maxTokens") == 64
         assert "maxCompletionTokens" not in out
 
+    @pytest.mark.parametrize("model", ["google.gemini-2.5-pro", "oci/google.gemini-2.5-flash"])
+    @pytest.mark.parametrize(("params", "expected"), [({"max_tokens": 64}, 64), ({}, DEFAULT_OCI_CHAT_MAX_TOKENS)])
+    def test_gemini_reasoning_models_keep_max_tokens(
+        self, model, params, expected, _flag_oci_gemini_as_reasoning_in_catalog
+    ):
+        """Regression: OCI's Gemini models ignore maxCompletionTokens and fall back to a default of
+        about 4K output tokens (verified against live OCI on 2026-09-28), so a reasoning flag in the
+        catalog must not move the limit off maxTokens."""
+        out = OCIChatConfig()._get_optional_params(OCIVendors.GENERIC, params, model=model)
+        assert out.get("maxTokens") == expected
+        assert "maxCompletionTokens" not in out
+
     def test_cohere_reasoning_model_keeps_max_tokens(self):
         cfg = OCIChatConfig()
         out = cfg._get_optional_params(