Merge f73ec1875c into f191e08d67

litellm/llms/anthropic/pass_through/messages/handler.py

@@ -70,10 +70,22 @@ def _responses_mode_is_lost_by_prefix_strip(
     )
 
 
+def _points_at_openai_backend(api_base: str | None) -> bool:
+    """Only real OpenAI hosts keep the Responses path; a custom api_base means a self-hosted
+    openai/-compatible backend (vLLM, llama.cpp, ...) that lacks the Responses API."""
+    if not api_base:
+        return True
+    from urllib.parse import urlparse
+
+    host = (urlparse(api_base if "://" in api_base else f"//{api_base}").hostname or "").lower()
+    return host == "api.openai.com" or host.endswith(".openai.com")
+
+
 def _should_route_to_responses_api(
     custom_llm_provider: str | None,
     requested_model: str | None = None,
     resolved_model: str | None = None,
+    api_base: str | None = None,
 ) -> bool:
     """Return True when the request should use the Responses API path.
 
@@ -83,7 +95,7 @@ def _should_route_to_responses_api(
     if litellm.use_chat_completions_url_for_anthropic_messages:
         return False
     if custom_llm_provider in _RESPONSES_API_PROVIDERS:
-        return True
+        return _points_at_openai_backend(api_base)
     if custom_llm_provider is None or requested_model is None or resolved_model is None:
         return False
     return _responses_mode_is_lost_by_prefix_strip(requested_model, resolved_model, custom_llm_provider)
@@ -581,7 +593,7 @@ def anthropic_messages_handler(
     if anthropic_messages_provider_config is None:
         # Route to Responses API for OpenAI / Azure, chat/completions for everything else.
         if kwargs.get("compaction") is None and _should_route_to_responses_api(
-            custom_llm_provider, original_model, model
+            custom_llm_provider, original_model, model, dynamic_api_base or api_base
         ):
             return LiteLLMMessagesToResponsesAPIHandler.anthropic_messages_handler(
                 max_tokens=max_tokens,

tests/unit/llms/anthropic/pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py

@@ -137,6 +137,50 @@ def test_anthropic_experimental_pass_through_messages_handler_dynamic_api_key_an
         assert mock_completion.call_args.kwargs["custom_key"] == "custom_value"
 
 
+@pytest.mark.parametrize(
+    "api_base, expected",
+    [
+        (None, True),
+        ("https://api.openai.com/v1", True),
+        ("api.openai.com", True),
+        ("HTTPS://API.OPENAI.COM/v1", True),
+        ("http://localhost:8000/v1", False),
+        ("http://vllm-host:8000/v1", False),
+        ("https://my-org.privatelink.openai.com/v1", True),
+        ("https://api.openai.com.evil.example/v1", False),
+        ("https://not-api.openai.com.internal/v1", False),
+    ],
+)
+def test_should_route_to_responses_api_considers_api_base_for_openai(api_base, expected):
+    """Regression test for #40780: openai/ with a custom api_base routes to chat/completions, real OpenAI hosts keep Responses."""
+    from litellm.llms.anthropic.pass_through.messages.handler import (
+        _should_route_to_responses_api,
+    )
+
+    assert _should_route_to_responses_api("openai", "openai/model", "model", api_base) is expected
+
+
+def test_openai_custom_api_base_forwards_messages_to_chat_completions():  # test-quality-ok: routing guard, asserts /v1/messages reaches litellm.completion with api_base forwarded; the routing decision itself is behaviorally covered by test_should_route_to_responses_api_considers_api_base_for_openai
+    """Regression test for #40780: the full handler forwards a custom-api_base openai/ request to litellm.completion."""
+    from litellm.llms.anthropic.pass_through.messages.handler import (
+        anthropic_messages_handler,
+    )
+
+    with patch("litellm.completion") as mock_completion:
+        try:
+            anthropic_messages_handler(
+                max_tokens=100,
+                messages=[{"role": "user", "content": "Hello, how are you?"}],
+                model="openai/my-local-model",
+                api_base="http://localhost:8000/v1",
+                api_key="sk-noauth",
+            )
+        except (ValueError, TypeError, AttributeError) as e:
+            print(f"Error: {e}")
+        mock_completion.assert_called_once()
+        assert mock_completion.call_args.kwargs["api_base"] == "http://localhost:8000/v1"
+
+
 @pytest.mark.asyncio
 async def test_anthropic_messages_sanitizes_empty_text_blocks_before_dispatch():
     """Regression test for #22930.  The unified /v1/messages path must