Merge bb9a71c285 into e52f05566d

litellm/__init__.py

@@ -673,6 +673,7 @@ wandb_models: Set = set(WANDB_MODELS)
 ovhcloud_models: Set = set()
 ovhcloud_embedding_models: Set = set()
 lemonade_models: Set = set()
+lm_studio_models: Set = set()  # mutable-ok: dynamic-discovery-only, like lemonade_models
 docker_model_runner_models: Set = set()
 amazon_nova_models: Set = set()
 stability_models: Set = set()
@@ -1191,6 +1192,7 @@ def _build_models_by_provider() -> dict:
         "wandb": wandb_models,
         "ovhcloud": ovhcloud_models | ovhcloud_embedding_models,
         "lemonade": lemonade_models,
+        "lm_studio": lm_studio_models,
         "clarifai": clarifai_models,
         "amazon_nova": amazon_nova_models,
         "stability": stability_models,

litellm/llms/lm_studio/chat/transformation.py

@@ -19,6 +19,25 @@ class LMStudioChatConfig(OpenAIGPTConfig):
         )  # LM Studio does not require an api key, but OpenAI client requires non-None value
         return api_base, dynamic_api_key
 
+    def get_models(self, api_key: str | None = None, api_base: str | None = None) -> list[str]:
+        """
+        Calls LM Studio's `/v1/models` endpoint and returns the list of models,
+        prefixed with "lm_studio/" (matching the OllamaModelInfo convention)
+        so discovered ids are directly callable without the caller having to
+        add the provider prefix themselves.
+
+        Reuses the same api_base/api_key resolution as chat completions
+        (LM_STUDIO_API_BASE / LM_STUDIO_API_KEY env vars, "fake-api-key"
+        fallback) instead of OpenAIGPTConfig's default of
+        https://api.openai.com, which would be wrong for a local/self-hosted
+        LM Studio server.
+        """
+        api_base, api_key = self._get_openai_compatible_provider_info(api_base=api_base, api_key=api_key)
+        models: Final = super().get_models(api_key=api_key, api_base=api_base)
+        return [  # mutable-ok: matches OllamaModelInfo.get_models' list-of-prefixed-ids contract
+            m if m.startswith("lm_studio/") else f"lm_studio/{m}" for m in models
+        ]
+
     def map_openai_params(
         self,
         non_default_params: dict,

litellm/utils.py

@@ -8664,24 +8664,31 @@ class ProviderConfigManager:
         model: str | None,
         provider: LlmProviders,
     ) -> BaseLLMModelInfo | None:
-        if LlmProviders.FIREWORKS_AI == provider:
-            return litellm.FireworksAIConfig()
-        elif LlmProviders.OPENAI == provider:
-            return litellm.OpenAIGPTConfig()
-        elif LlmProviders.GEMINI == provider:
-            return litellm.GeminiModelInfo()
+        # Providers whose BaseLLMModelInfo needs no local import and no `model`
+        # arg: a single dict lookup instead of N `elif` branches keeps this
+        # dispatcher's cyclomatic complexity from growing with every provider
+        # added here (see the elif chain below for providers that DO need a
+        # local import or the `model` arg).
+        simple_provider_configs: Final[
+            dict[LlmProviders, Callable[[], BaseLLMModelInfo]]
+        ] = {  # mutable-ok: built once per call, read-only lookup table replacing 9 trivial elif branches
+            LlmProviders.FIREWORKS_AI: litellm.FireworksAIConfig,
+            LlmProviders.OPENAI: litellm.OpenAIGPTConfig,
+            LlmProviders.GEMINI: litellm.GeminiModelInfo,
+            LlmProviders.LITELLM_PROXY: litellm.LiteLLMProxyChatConfig,
+            LlmProviders.TOPAZ: litellm.TopazModelInfo,
+            LlmProviders.ANTHROPIC: litellm.AnthropicModelInfo,
+            LlmProviders.XAI: litellm.XAIModelInfo,
+            LlmProviders.LEMONADE: litellm.LemonadeChatConfig,
+            LlmProviders.CLARIFAI: litellm.ClarifaiConfig,
+            LlmProviders.LM_STUDIO: litellm.LMStudioChatConfig,
+        }
+        if provider in simple_provider_configs:
+            return simple_provider_configs[provider]()
         elif LlmProviders.VERTEX_AI == provider:
             from litellm.llms.vertex_ai.common_utils import VertexAIModelInfo
 
             return VertexAIModelInfo()
-        elif LlmProviders.LITELLM_PROXY == provider:
-            return litellm.LiteLLMProxyChatConfig()
-        elif LlmProviders.TOPAZ == provider:
-            return litellm.TopazModelInfo()
-        elif LlmProviders.ANTHROPIC == provider:
-            return litellm.AnthropicModelInfo()
-        elif LlmProviders.XAI == provider:
-            return litellm.XAIModelInfo()
         elif LlmProviders.OLLAMA == provider or LlmProviders.OLLAMA_CHAT == provider:
             # Dynamic model listing for Ollama server
             from litellm.llms.ollama.common_utils import OllamaModelInfo
@@ -8693,10 +8700,6 @@ class ProviderConfigManager:
             )
 
             return VLLMModelInfo()
-        elif LlmProviders.LEMONADE == provider:
-            return litellm.LemonadeChatConfig()
-        elif LlmProviders.CLARIFAI == provider:
-            return litellm.ClarifaiConfig()
         elif LlmProviders.BEDROCK == provider:
             from litellm.llms.bedrock.common_utils import BedrockModelInfo
 

tests/test_litellm/test_utils.py

@@ -2887,6 +2887,58 @@ class TestGetValidModelsWithCLI:
             assert headers.get("Authorization") == "Bearer sk-test-cli-key-123"
 
 
+class TestGetValidModelsLMStudio:
+    """Test get_valid_models(check_provider_endpoint=True) for lm_studio.
+
+    get_provider_model_info() previously had no LM_STUDIO branch, so
+    discovery silently returned an empty list regardless of
+    check_provider_endpoint; and lm_studio was missing from
+    models_by_provider, so the proxy's wildcard-model expansion
+    (get_provider_models) bailed out before ever calling get_valid_models.
+    """
+
+    def test_get_valid_models_lm_studio_discovery(self):
+        """Discovery hits LM Studio's own /v1/models, not OpenAI's default api_base, and prefixes results with lm_studio/."""
+        mock_response = MagicMock()
+        mock_response.status_code = 200
+        mock_response.json.return_value = {
+            "data": [
+                {"id": "qwen/qwen3-a3b", "object": "model"},
+                {"id": "lm_studio/already-prefixed", "object": "model"},
+            ]
+        }
+
+        with patch.object(
+            litellm.module_level_client, "get", return_value=mock_response
+        ) as mock_get:
+            result = litellm.get_valid_models(
+                check_provider_endpoint=True,
+                custom_llm_provider="lm_studio",
+                api_key="lm-studio-key",
+                api_base="http://my-lm-studio-host:1234",
+            )
+
+            assert isinstance(result, list)
+            assert result == [
+                "lm_studio/qwen/qwen3-a3b",
+                "lm_studio/already-prefixed",
+            ]
+
+            mock_get.assert_called_once()
+            _, call_kwargs = mock_get.call_args
+
+            # Must hit the passed-in LM Studio host, not OpenAIGPTConfig's
+            # default of https://api.openai.com
+            assert call_kwargs["url"] == "http://my-lm-studio-host:1234/v1/models"
+            assert call_kwargs["headers"]["Authorization"] == "Bearer lm-studio-key"
+
+    def test_lm_studio_in_models_by_provider(self):
+        """lm_studio must be a key in models_by_provider or the proxy's
+        wildcard-model expansion (get_provider_models) bails out before
+        ever reaching get_valid_models/check_provider_endpoint."""
+        assert "lm_studio" in litellm.models_by_provider
+
+
 class TestIsCachedMessage:
     """Test is_cached_message function for context caching detection.