Merge adf523d65b into 5ddcc45b3a

litellm/utils.py

@@ -10012,6 +10012,19 @@ class ProviderConfigManager:
             if VertexAIPartnerModels.is_vertex_partner_model(model):
                 return None
 
+            #########################################################
+            # Model Garden publisher models served via the OpenAI-compatible
+            # route (e.g. xai/grok-*) are not Vertex Partner "raw" models, but
+            # they still go through the litellm.completion() adapter (which
+            # dispatches them to the Model Garden OpenAI-compatible handler),
+            # never through the Google Gen AI config. Without this case the
+            # unified Gemini-style endpoint POST /v1beta/models/{model}:generateContent
+            # would treat e.g. vertex_ai/xai/grok-* as a native Gemini model
+            # and fail with a protocol mismatch.
+            #########################################################
+            if model.startswith("xai/"):
+                return None
+
             #########################################################
             # If the model is not a Vertex Partner model, return the Vertex AI Google Gen AI Config
             # This is for Vertex `gemini` models

tests/unit/test_utils.py

@@ -6447,6 +6447,53 @@ def test_calculate_max_parallel_requests_precedence(
     )
 
 
+def test_get_provider_google_genai_generate_content_config_vertex_partner_models():
+    """
+    https://github.com/BerriAI/litellm/issues/41835
+    Vertex Grok (vertex_ai/xai/grok-*) is rejected on /v1beta/models/{model}:generateContent
+    because get_provider_google_genai_generate_content_config() only falls back to the
+    litellm.completion() adapter for Vertex Partner models, and xai/ was missing from
+    PartnerModelPrefixes. Ensure xai/grok-* now returns None (i.e. the generateContent
+    endpoint routes it through the chat-completions adapter like other Model Garden
+    OpenAI-compatible partners), while gemini models still get the native config.
+    """
+    from litellm.types.utils import LlmProviders
+
+    # Grok via Model Garden -> through the completion adapter (None)
+    assert (
+        ProviderConfigManager.get_provider_google_genai_generate_content_config(
+            model="xai/grok-4.6", provider=LlmProviders.VERTEX_AI
+        )
+        is None
+    )
+    assert (
+        ProviderConfigManager.get_provider_google_genai_generate_content_config(
+            model="xai/grok-4.1-fast-non-reasoning", provider=LlmProviders.VERTEX_AI
+        )
+        is None
+    )
+
+    # Other existing partner models are unaffected
+    assert (
+        ProviderConfigManager.get_provider_google_genai_generate_content_config(
+            model="claude-3-5-sonnet-20241022", provider=LlmProviders.VERTEX_AI
+        )
+        is None
+    )
+
+    # Native Gemini models still get the Vertex Google Gen AI config
+    from litellm.llms.vertex_ai.google_genai.transformation import (
+        VertexAIGoogleGenAIConfig,
+    )
+
+    assert isinstance(
+        ProviderConfigManager.get_provider_google_genai_generate_content_config(
+            model="gemini-1.5-pro", provider=LlmProviders.VERTEX_AI
+        ),
+        VertexAIGoogleGenAIConfig,
+    )
+
+
 class _NamedStream(io.BytesIO):
     def __init__(self, name: str | int) -> None:
         super().__init__(b"%PDF-1.4 secret document body")