fix: apply same credential caching fix to VertexAIGemmaModels and VertexAIModelGardenModels

Same bug as VertexAIPartnerModels: both classes had `pass` in __init__
instead of `super().__init__()`, and created throwaway VertexLLM()
instances per request instead of using self._ensure_access_token().

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

litellm/llms/vertex_ai/vertex_gemma_models/main.py

@@ -31,7 +31,7 @@ from ..vertex_llm_base import VertexBase
 
 class VertexAIGemmaModels(VertexBase):
     def __init__(self) -> None:
-        pass
+        super().__init__()
 
     def completion(
         self,
@@ -62,9 +62,6 @@ class VertexAIGemmaModels(VertexBase):
         try:
             import vertexai
 
-            from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
-                VertexLLM,
-            )
             from litellm.llms.vertex_ai.vertex_gemma_models.transformation import (
                 VertexGemmaConfig,
             )
@@ -83,9 +80,8 @@ class VertexAIGemmaModels(VertexBase):
             )
         try:
             model = get_vertex_base_model_name(model=model)
-            vertex_httpx_logic = VertexLLM()
 
-            access_token, project_id = vertex_httpx_logic._ensure_access_token(
+            access_token, project_id = self._ensure_access_token(
                 credentials=vertex_credentials,
                 project_id=vertex_project,
                 custom_llm_provider="vertex_ai",

litellm/llms/vertex_ai/vertex_model_garden/main.py

@@ -41,7 +41,7 @@ def create_vertex_url(
 
 class VertexAIModelGardenModels(VertexBase):
     def __init__(self) -> None:
-        pass
+        super().__init__()
 
     def completion(
         self,
@@ -73,9 +73,6 @@ class VertexAIModelGardenModels(VertexBase):
             import vertexai
 
             from litellm.llms.openai_like.chat.handler import OpenAILikeChatHandler
-            from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
-                VertexLLM,
-            )
         except Exception as e:
             raise VertexAIError(
                 status_code=400,
@@ -91,9 +88,8 @@ class VertexAIModelGardenModels(VertexBase):
             )
         try:
             model = get_vertex_base_model_name(model=model)
-            vertex_httpx_logic = VertexLLM()
 
-            access_token, project_id = vertex_httpx_logic._ensure_access_token(
+            access_token, project_id = self._ensure_access_token(
                 credentials=vertex_credentials,
                 project_id=vertex_project,
                 custom_llm_provider="vertex_ai",

tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/test_partner_models_credential_reuse.py

@@ -1,40 +1,53 @@
 """
-Test that VertexAIPartnerModels reuses cached credentials from VertexBase
-instead of creating a new VertexLLM instance on every request.
+Test that VertexBase subclasses (PartnerModels, Gemma, ModelGarden) reuse
+cached credentials instead of creating a new VertexLLM instance on every request.
 """
 
 import sys
 from unittest.mock import MagicMock, patch
 
+import pytest
+
 from litellm.llms.vertex_ai.vertex_ai_partner_models.main import (
     VertexAIPartnerModels,
 )
+from litellm.llms.vertex_ai.vertex_gemma_models.main import VertexAIGemmaModels
+from litellm.llms.vertex_ai.vertex_model_garden.main import VertexAIModelGardenModels
+
+
+def _mock_vertexai():
+    """Return a MagicMock that satisfies the vertexai import guards."""
+    m = MagicMock()
+    m.preview = MagicMock()
+    m.preview.language_models = MagicMock()
+    return m
+
+
+class TestVertexBaseSubclassInit:
+    """All VertexBase subclasses must call super().__init__() so that
+    the credential cache is initialized."""
+
+    @pytest.mark.parametrize(
+        "cls",
+        [VertexAIPartnerModels, VertexAIGemmaModels, VertexAIModelGardenModels],
+        ids=["PartnerModels", "Gemma", "ModelGarden"],
+    )
+    def test_init_calls_super(self, cls):
+        instance = cls()
+        assert hasattr(instance, "_credentials_project_mapping")
+        assert isinstance(instance._credentials_project_mapping, dict)
+        assert hasattr(instance, "access_token")
+        assert hasattr(instance, "project_id")
 
 
 class TestPartnerModelsCredentialReuse:
-    def test_init_calls_super(self):
-        """VertexAIPartnerModels.__init__ must call super().__init__() so that
-        the VertexBase credential cache is initialized."""
-        partner = VertexAIPartnerModels()
-        # These attributes are set by VertexBase.__init__
-        assert hasattr(partner, "_credentials_project_mapping")
-        assert isinstance(partner._credentials_project_mapping, dict)
-        assert hasattr(partner, "access_token")
-        assert hasattr(partner, "project_id")
-
     def test_completion_uses_self_ensure_access_token(self):
         """completion() should call self._ensure_access_token, not create a
-        throwaway VertexLLM instance. This ensures the credential cache on the
-        singleton is reused across calls."""
+        throwaway VertexLLM instance."""
         partner = VertexAIPartnerModels()
 
-        # Mock vertexai import and the completion handler
-        mock_vertexai = MagicMock()
-        mock_vertexai.preview = MagicMock()
-        mock_vertexai.preview.language_models = MagicMock()
-
         with (
-            patch.dict(sys.modules, {"vertexai": mock_vertexai}),
+            patch.dict(sys.modules, {"vertexai": _mock_vertexai()}),
             patch.object(
                 partner,
                 "_ensure_access_token",
@@ -64,7 +77,6 @@ class TestPartnerModelsCredentialReuse:
                 vertex_credentials='{"type": "service_account"}',
             )
 
-            # _ensure_access_token should have been called on self
             mock_ensure.assert_called_once_with(
                 credentials='{"type": "service_account"}',
                 project_id="test-project",
@@ -72,8 +84,7 @@ class TestPartnerModelsCredentialReuse:
             )
 
     def test_credential_cache_shared_across_calls(self):
-        """Two successive completion() calls should hit load_auth only once,
-        proving the credential cache on the VertexAIPartnerModels instance works."""
+        """Two successive completion() calls should hit load_auth only once."""
         partner = VertexAIPartnerModels()
 
         mock_creds = MagicMock()
@@ -82,12 +93,8 @@ class TestPartnerModelsCredentialReuse:
         mock_creds.project_id = "proj"
         mock_creds.quota_project_id = "proj"
 
-        mock_vertexai = MagicMock()
-        mock_vertexai.preview = MagicMock()
-        mock_vertexai.preview.language_models = MagicMock()
-
         with (
-            patch.dict(sys.modules, {"vertexai": mock_vertexai}),
+            patch.dict(sys.modules, {"vertexai": _mock_vertexai()}),
             patch.object(
                 partner, "load_auth", return_value=(mock_creds, "proj")
             ) as mock_load,
@@ -118,5 +125,96 @@ class TestPartnerModelsCredentialReuse:
             partner.completion(**common_kwargs)
             partner.completion(**common_kwargs)
 
-            # load_auth should only be called once — second call uses cache
             assert mock_load.call_count == 1
+
+
+class TestGemmaModelsCredentialReuse:
+    def test_completion_uses_self_ensure_access_token(self):
+        """completion() should call self._ensure_access_token, not create a
+        throwaway VertexLLM instance."""
+        gemma = VertexAIGemmaModels()
+
+        mock_gemma_config = MagicMock()
+        mock_gemma_config.return_value.completion.return_value = "response"
+
+        with (
+            patch.dict(sys.modules, {"vertexai": _mock_vertexai()}),
+            patch.object(
+                gemma,
+                "_ensure_access_token",
+                return_value=("cached-token", "test-project"),
+            ) as mock_ensure,
+            patch(
+                "litellm.llms.vertex_ai.vertex_gemma_models.transformation.VertexGemmaConfig",
+                mock_gemma_config,
+            ),
+        ):
+            gemma.completion(
+                model="gemma/gemma-3-12b-it-1234567890",
+                messages=[{"role": "user", "content": "hello"}],
+                model_response=MagicMock(),
+                print_verbose=lambda *a, **kw: None,
+                encoding=MagicMock(),
+                logging_obj=MagicMock(),
+                api_base="https://123.us-central1-1.prediction.vertexai.goog/v1/projects/proj/locations/us-central1/endpoints/456:predict",
+                optional_params={},
+                custom_prompt_dict={},
+                headers=None,
+                timeout=30.0,
+                litellm_params={},
+                vertex_project="test-project",
+                vertex_location="us-central1",
+                vertex_credentials='{"type": "service_account"}',
+            )
+
+            mock_ensure.assert_called_once_with(
+                credentials='{"type": "service_account"}',
+                project_id="test-project",
+                custom_llm_provider="vertex_ai",
+            )
+
+
+class TestModelGardenCredentialReuse:
+    def test_completion_uses_self_ensure_access_token(self):
+        """completion() should call self._ensure_access_token, not create a
+        throwaway VertexLLM instance."""
+        garden = VertexAIModelGardenModels()
+
+        mock_handler = MagicMock()
+        mock_handler.return_value.completion.return_value = "response"
+
+        with (
+            patch.dict(sys.modules, {"vertexai": _mock_vertexai()}),
+            patch.object(
+                garden,
+                "_ensure_access_token",
+                return_value=("cached-token", "test-project"),
+            ) as mock_ensure,
+            patch(
+                "litellm.llms.openai_like.chat.handler.OpenAILikeChatHandler",
+                mock_handler,
+            ),
+        ):
+            garden.completion(
+                model="openai/5464397967697903616",
+                messages=[{"role": "user", "content": "hello"}],
+                model_response=MagicMock(),
+                print_verbose=lambda *a, **kw: None,
+                encoding=MagicMock(),
+                logging_obj=MagicMock(),
+                api_base=None,
+                optional_params={},
+                custom_prompt_dict={},
+                headers=None,
+                timeout=30.0,
+                litellm_params={},
+                vertex_project="test-project",
+                vertex_location="us-central1",
+                vertex_credentials='{"type": "service_account"}',
+            )
+
+            mock_ensure.assert_called_once_with(
+                credentials='{"type": "service_account"}',
+                project_id="test-project",
+                custom_llm_provider="vertex_ai",
+            )