Merge 31ae093cbf into 883b722fd2

litellm/main.py

@@ -1065,7 +1065,7 @@ def responses_api_bridge_check(
     try:
         model_info = cast(
             dict,
-            _get_model_info_helper(model=model, custom_llm_provider=custom_llm_provider),
+            _get_model_info_helper(model=model, custom_llm_provider=custom_llm_provider, api_base=api_base),
         )
         if model_info.get("mode") is None and model.startswith("responses/"):
             model = model.replace("responses/", "")
@@ -4325,7 +4325,7 @@ def _complete_ollama(ctx: _CompletionDispatchContext) -> _CompletionDispatchResu
     stream: Final = ctx.stream
     timeout: Final = ctx.timeout
 
-    api_base = litellm.api_base or api_base or get_secret("OLLAMA_API_BASE") or "http://localhost:11434"
+    api_base = api_base or litellm.api_base or get_secret("OLLAMA_API_BASE") or "http://localhost:11434"
     if api_key is not None and "Authorization" not in headers:
         headers["Authorization"] = f"Bearer {api_key}"
 
@@ -4365,7 +4365,7 @@ def _complete_ollama_chat(ctx: _CompletionDispatchContext) -> _CompletionDispatc
     stream: Final = ctx.stream
     timeout: Final = ctx.timeout
 
-    api_base = litellm.api_base or api_base or get_secret("OLLAMA_API_BASE") or "http://localhost:11434"
+    api_base = api_base or litellm.api_base or get_secret("OLLAMA_API_BASE") or "http://localhost:11434"
 
     api_key = api_key or litellm.ollama_key or os.environ.get("OLLAMA_API_KEY") or litellm.api_key
     if api_key is not None and "Authorization" not in headers:
@@ -6744,7 +6744,7 @@ def embedding(
                 api_key=api_key,
             )
         elif custom_llm_provider == "ollama":
-            api_base = litellm.api_base or api_base or get_secret_str("OLLAMA_API_BASE") or "http://localhost:11434"
+            api_base = api_base or litellm.api_base or get_secret_str("OLLAMA_API_BASE") or "http://localhost:11434"
 
             if isinstance(input, str):
                 input = [input]

tests/test_litellm/test_main.py

@@ -815,6 +815,96 @@ def test_responses_api_bridge_check_gpt_5_5_tools_plus_reasoning_routes_to_respo
     assert model_info.get("mode") == "responses"
 
 
+_OLLAMA_SHOW_RESPONSE: Final = {
+    "model_info": {"llama.context_length": 8192},
+    "capabilities": ["completion"],
+}
+
+_OLLAMA_GENERATE_RESPONSE: Final = {
+    "model": "llama3-custom",
+    "response": "hello from ollama",
+    "done": True,
+    "done_reason": "stop",
+    "prompt_eval_count": 5,
+    "eval_count": 4,
+}
+
+_OLLAMA_CHAT_RESPONSE: Final = {
+    "model": "llama3-custom",
+    "message": {"role": "assistant", "content": "hello from ollama"},
+    "done": True,
+    "done_reason": "stop",
+    "prompt_eval_count": 5,
+    "eval_count": 4,
+}
+
+
+def test_responses_api_bridge_check_forwards_api_base_to_model_info_lookup(respx_mock: respx.MockRouter):
+    """Regression test for https://github.com/BerriAI/litellm/issues/37041 -- the bridge
+    check's model-info lookup must hit the request's api_base, not fall back to the
+    provider default (localhost:11434 for ollama)."""
+    from litellm.main import responses_api_bridge_check
+
+    show_route = respx_mock.post("http://my-host:30000/api/show").respond(json=_OLLAMA_SHOW_RESPONSE)
+
+    litellm.get_model_info.cache_clear()
+    model_info, model = responses_api_bridge_check(
+        model="llama3-custom",
+        custom_llm_provider="ollama",
+        api_base="http://my-host:30000",
+    )
+
+    assert show_route.called
+    assert model == "llama3-custom"
+    assert model_info["max_tokens"] == 8192
+
+
+@pytest.mark.parametrize(
+    "model, completion_path, completion_response",
+    [
+        ("ollama/llama3-custom", "/api/generate", _OLLAMA_GENERATE_RESPONSE),
+        ("ollama_chat/llama3-custom", "/api/chat", _OLLAMA_CHAT_RESPONSE),
+    ],
+)
+def test_ollama_completion_explicit_api_base_overrides_global(
+    model, completion_path, completion_response, monkeypatch, respx_mock: respx.MockRouter
+):
+    """Regression test for https://github.com/BerriAI/litellm/issues/26170 -- the explicit
+    api_base kwarg must win over the litellm.api_base global, matching the openai provider."""
+    monkeypatch.setattr(litellm, "api_base", "https://unrelated-global.example.com")
+    respx_mock.post("http://my-host:30000/api/show").respond(json=_OLLAMA_SHOW_RESPONSE)
+    completion_route = respx_mock.post(f"http://my-host:30000{completion_path}").respond(json=completion_response)
+
+    litellm.get_model_info.cache_clear()
+    response = litellm.completion(
+        model=model,
+        messages=[{"role": "user", "content": "hi"}],
+        api_base="http://my-host:30000",
+    )
+
+    assert completion_route.called
+    assert response.choices[0].message.content == "hello from ollama"
+
+
+def test_ollama_embedding_explicit_api_base_overrides_global(monkeypatch, respx_mock: respx.MockRouter):
+    """Regression test for https://github.com/BerriAI/litellm/issues/26170 (embedding path)."""
+    monkeypatch.setattr(litellm, "api_base", "https://unrelated-global.example.com")
+    respx_mock.post("http://my-host:30000/api/show").respond(json=_OLLAMA_SHOW_RESPONSE)
+    embed_route = respx_mock.post("http://my-host:30000/api/embed").respond(
+        json={"model": "qwen3-embedding:0.6b", "embeddings": [[0.1, 0.2, 0.3]], "prompt_eval_count": 2}
+    )
+
+    litellm.get_model_info.cache_clear()
+    response = litellm.embedding(
+        model="ollama/qwen3-embedding:0.6b",
+        input="hello",
+        api_base="http://my-host:30000",
+    )
+
+    assert embed_route.called
+    assert response.data[0]["embedding"] == [0.1, 0.2, 0.3]
+
+
 def test_responses_api_bridge_check_azure_gpt_5_4_tools_plus_reasoning_routes_to_responses():
     """Azure gpt-5.4 with both tools and reasoning_effort should route to Responses API."""
     from litellm.main import responses_api_bridge_check