fix(vertex_ai): preserve messages and tools on fallback, support bare models and sync transform

litellm/llms/anthropic/messages/adapters/transformation.py

@@ -740,6 +740,13 @@ class LiteLLMAnthropicMessagesAdapter:
         if "gemini" in model.lower():
             from litellm.utils import is_explicitly_disabled_factory
 
+            if "/" not in model:
+                if (
+                    is_explicitly_disabled_factory(f"gemini/{model}", None, "supports_prompt_caching")
+                    or is_explicitly_disabled_factory(f"vertex_ai/{model}", None, "supports_prompt_caching")
+                ):
+                    return False
+
             if is_explicitly_disabled_factory(model, None, "supports_prompt_caching"):
                 return False
             return True

litellm/llms/vertex_ai/context_caching/vertex_ai_context_caching.py

@@ -507,9 +507,9 @@ class ContextCachingEndpoints(VertexBase):
         except _CachedContentsNotServedError:
             verbose_logger.debug("Gemini context caching: %s answers 404. Skipping context caching.", url)
             return messages, optional_params, None
-        for tool_param in _TOOL_PARAMS:
-            optional_params.pop(tool_param, None)
         if google_cache_name:
+            for tool_param in _TOOL_PARAMS:
+                optional_params.pop(tool_param, None)
             return non_cached_messages, optional_params, google_cache_name
 
         ## TRANSFORM REQUEST
@@ -571,6 +571,8 @@ class ContextCachingEndpoints(VertexBase):
             raw_response_cached.get("name"),
             raw_response_cached.get("expireTime"),
         )
+        for tool_param in _TOOL_PARAMS:
+            optional_params.pop(tool_param, None)
         return (
             non_cached_messages,
             optional_params,
@@ -738,6 +740,10 @@ class ContextCachingEndpoints(VertexBase):
         except _CachedContentsNotServedError:
             verbose_logger.debug("Gemini context caching: %s answers 404. Skipping context caching.", url)
             return messages, optional_params, None
+
+        if cache_name is None:
+            return messages, optional_params, None
+
         for tool_param in _TOOL_PARAMS:
             optional_params.pop(tool_param, None)
         return non_cached_messages, optional_params, cache_name

litellm/llms/vertex_ai/gemini/transformation.py

@@ -1262,6 +1262,11 @@ def sync_transform_request_body(
 
     context_caching_endpoints: Final = ContextCachingEndpoints()
 
+    if context_caching:
+        model_info = litellm_params.get("model_info") or {}
+        if model_info.get("supports_prompt_caching") is False:
+            context_caching = False
+
     (
         messages,
         optional_params,

tests/unit/llms/vertex_ai/context_caching/test_vertex_ai_explicit_caching_unsupported.py

@@ -11,20 +11,23 @@ from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
 from litellm.llms.vertex_ai.context_caching.vertex_ai_context_caching import (
     ContextCachingEndpoints,
 )
+from litellm.llms.vertex_ai.gemini.transformation import (
+    sync_transform_request_body,
+)
 
 
-def test_target_consumes_cache_control_skips_when_model_disables_prompt_caching(monkeypatch):
+def test_target_consumes_cache_control_skips_when_model_disables_prompt_caching(local_model_cost_map):
     """LiteLLMAnthropicMessagesAdapter should not consume cache_control for models explicitly disabling prompt caching."""
-    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
     assert LiteLLMAnthropicMessagesAdapter.target_consumes_cache_control("claude-3-5-sonnet-20241022") is True
     assert LiteLLMAnthropicMessagesAdapter.target_consumes_cache_control("gemini/gemini-2.5-flash") is True
     assert LiteLLMAnthropicMessagesAdapter.target_consumes_cache_control("gemini/gemini-4-argon") is False
     assert LiteLLMAnthropicMessagesAdapter.target_consumes_cache_control("vertex_ai/gemini-4-argon") is False
+    # Bare model name without provider prefix should also be recognized
+    assert LiteLLMAnthropicMessagesAdapter.target_consumes_cache_control("gemini-4-argon") is False
 
 
-def test_check_and_create_cache_skips_when_model_disables_prompt_caching(monkeypatch):
+def test_check_and_create_cache_skips_when_model_disables_prompt_caching(local_model_cost_map):
     """ContextCachingEndpoints should skip explicit cache creation when model has supports_prompt_caching: false."""
-    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
     endpoints = ContextCachingEndpoints()
     messages = [
         {
@@ -55,14 +58,14 @@ def test_check_and_create_cache_skips_when_model_disables_prompt_caching(monkeyp
         vertex_auth_header=None,
     )
     assert cache_id is None
-    # Tools should remain in params and not be popped
+    # All messages and tools should remain intact
+    assert len(res_messages) == len(messages)
     assert "tools" in res_params
 
 
 @pytest.mark.asyncio
-async def test_async_check_and_create_cache_fallback_on_400_explicit_caching_unsupported(monkeypatch):
+async def test_async_check_and_create_cache_fallback_on_400_explicit_caching_unsupported(local_model_cost_map):
     """When Vertex AI returns 400 'does not support explicit context caching', gracefully fallback to non-cached request."""
-    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
     endpoints = ContextCachingEndpoints()
     endpoints.async_check_cache = AsyncMock(return_value=None)
     messages = [
@@ -109,11 +112,13 @@ async def test_async_check_and_create_cache_fallback_on_400_explicit_caching_uns
     )
     assert cache_id is None
     assert mock_client.post.called
+    # Critical checks: messages must not be dropped, tools must not be stripped!
+    assert len(res_messages) == len(messages)
+    assert "tools" in res_params
 
 
-def test_sync_check_and_create_cache_fallback_on_400_explicit_caching_unsupported(monkeypatch):
+def test_sync_check_and_create_cache_fallback_on_400_explicit_caching_unsupported(local_model_cost_map):
     """Sync check_and_create_cache fallback on 400 explicit caching unsupported."""
-    monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
     endpoints = ContextCachingEndpoints()
     endpoints.check_cache = MagicMock(return_value=None)
     messages = [
@@ -160,3 +165,38 @@ def test_sync_check_and_create_cache_fallback_on_400_explicit_caching_unsupporte
     )
     assert cache_id is None
     assert mock_client.post.called
+    # Critical checks: messages must not be dropped, tools must not be stripped!
+    assert len(res_messages) == len(messages)
+    assert "tools" in res_params
+
+
+def test_sync_transform_request_body_skips_caching_on_model_info_override(monkeypatch):
+    """sync_transform_request_body should respect supports_prompt_caching: false in model_info."""
+    mock_endpoints = MagicMock()
+    mock_endpoints.check_and_create_cache.return_value = ([], {}, None)
+    monkeypatch.setattr(
+        "litellm.llms.vertex_ai.context_caching.vertex_ai_context_caching.ContextCachingEndpoints",
+        MagicMock(return_value=mock_endpoints),
+    )
+
+    messages = [{"role": "user", "content": "hello"}]
+    sync_transform_request_body(
+        messages=messages,
+        model="gemini-2.5-flash",
+        custom_llm_provider="gemini",
+        litellm_params={"model_info": {"supports_prompt_caching": False}},
+        optional_params={},
+        gemini_api_key="test-key",
+        api_base=None,
+        client=None,
+        timeout=None,
+        extra_headers=None,
+        logging_obj=MagicMock(),
+        vertex_project=None,
+        vertex_location=None,
+        vertex_auth_header=None,
+    )
+    # context_caching=False must be passed to check_and_create_cache
+    assert mock_endpoints.check_and_create_cache.called
+    _, kwargs = mock_endpoints.check_and_create_cache.call_args
+    assert kwargs.get("context_caching") is False