From 843189740bbc7e30bad1449c8cf9d810a6afb330 Mon Sep 17 00:00:00 2001
From: Rainmemery <152621457+Rainmemery@users.noreply.github.com>
Date: Mon, 21 Sep 2026 18:33:05 +0000
Subject: [PATCH] fix(vertex_ai): route xai/grok through completion adapter on
 generateContent

POST /v1beta/models/{model}:generateContent treated vertex_ai/xai/grok-*
as a native Gemini model: get_provider_google_genai_generate_content_config()
only falls back to the litellm.completion() adapter for Vertex Partner
models, and PartnerModelPrefixes has no xai/ entry. Grok is dispatched to
VertexAIGoogleGenAIConfig and fails with a protocol mismatch.

xai/ publisher models are Model Garden OpenAI-compatible, matching how
get_vertex_ai_model_route() already serves them (MODEL_GARDEN). Route them
through the completion adapter on the unified Gemini-style endpoint, the
same way openai/gpt-oss-, moonshotai/, minimaxai/, etc. are handled.

Signed-off-by: Rainmemery <152621457+Rainmemery@users.noreply.github.com>
---
 litellm/utils.py                 | 13 +++++++++
 tests/test_litellm/test_utils.py | 46 ++++++++++++++++++++++++++++++++
 2 files changed, 59 insertions(+)

diff --git a/litellm/utils.py b/litellm/utils.py
index 252bc691301..b4f8e5b0d86 100644
--- a/litellm/utils.py
+++ b/litellm/utils.py
@@ -9767,6 +9767,19 @@ 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
diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py
index 7a7d5d44e95..24900c27877 100644
--- a/tests/test_litellm/test_utils.py
+++ b/tests/test_litellm/test_utils.py
@@ -5796,3 +5796,49 @@ def test_calculate_max_parallel_requests_precedence(
         )
         == expected
     )
+
+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,
+    )