feat: Add enable_model_discovery to enable/disable model discovery on startup

docs/docs/providers/inference/remote_vllm.mdx

@@ -20,6 +20,7 @@ Remote vLLM inference provider for connecting to vLLM servers.
 | `api_token` | `str \| None` | No | fake | The API token |
 | `tls_verify` | `bool \| str` | No | True | Whether to verify TLS certificates. Can be a boolean or a path to a CA certificate file. |
 | `refresh_models` | `<class 'bool'>` | No | False | Whether to refresh models periodically |
+| `allow_listing_models` | `<class 'bool'>` | No | True | Whether to allow listing models from the vLLM server |
 
 ## Sample Configuration
 
@@ -28,4 +29,5 @@ url: ${env.VLLM_URL:=}
 max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
 api_token: ${env.VLLM_API_TOKEN:=fake}
 tls_verify: ${env.VLLM_TLS_VERIFY:=true}
+allow_listing_models: ${env.VLLM_ALLOW_LISTING_MODELS:=true}
 ```

llama_stack/core/routing_tables/models.py

@@ -43,6 +43,12 @@ async def refresh(self) -> None:
             await self.update_registered_models(provider_id, models)
 
     async def list_models(self) -> ListModelsResponse:
+        # Check if providers allow listing models before returning models
+        for provider_id, provider in self.impls_by_provider_id.items():
+            allow_listing_models = await provider.allow_listing_models()
+            logger.debug(f"Provider {provider_id}: allow_listing_models={allow_listing_models}")
+            if not allow_listing_models:
+                logger.debug(f"Provider {provider_id} has allow_listing_models disabled")
         return ListModelsResponse(data=await self.get_all_with_type("model"))
 
     async def openai_list_models(self) -> OpenAIListModelsResponse:

llama_stack/distributions/ci-tests/run.yaml

@@ -31,6 +31,7 @@ providers:
       max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
       api_token: ${env.VLLM_API_TOKEN:=fake}
       tls_verify: ${env.VLLM_TLS_VERIFY:=true}
+      allow_listing_models: ${env.VLLM_ALLOW_LISTING_MODELS:=true}
   - provider_id: ${env.TGI_URL:+tgi}
     provider_type: remote::tgi
     config:

llama_stack/distributions/postgres-demo/run.yaml

@@ -16,6 +16,7 @@ providers:
       max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
       api_token: ${env.VLLM_API_TOKEN:=fake}
       tls_verify: ${env.VLLM_TLS_VERIFY:=true}
+      allow_listing_models: ${env.VLLM_ALLOW_LISTING_MODELS:=true}
   - provider_id: sentence-transformers
     provider_type: inline::sentence-transformers
   vector_io:

llama_stack/distributions/starter-gpu/run.yaml

@@ -31,6 +31,7 @@ providers:
       max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
       api_token: ${env.VLLM_API_TOKEN:=fake}
       tls_verify: ${env.VLLM_TLS_VERIFY:=true}
+      allow_listing_models: ${env.VLLM_ALLOW_LISTING_MODELS:=true}
   - provider_id: ${env.TGI_URL:+tgi}
     provider_type: remote::tgi
     config:

llama_stack/distributions/starter/run.yaml

@@ -31,6 +31,7 @@ providers:
       max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
       api_token: ${env.VLLM_API_TOKEN:=fake}
       tls_verify: ${env.VLLM_TLS_VERIFY:=true}
+      allow_listing_models: ${env.VLLM_ALLOW_LISTING_MODELS:=true}
   - provider_id: ${env.TGI_URL:+tgi}
     provider_type: remote::tgi
     config:

llama_stack/providers/inline/inference/meta_reference/inference.py

@@ -71,6 +71,9 @@ async def openai_completion(self, *args, **kwargs):
     async def should_refresh_models(self) -> bool:
         return False
 
+    async def allow_listing_models(self) -> bool:
+        return True
+
     async def list_models(self) -> list[Model] | None:
         return None
 

llama_stack/providers/inline/inference/sentence_transformers/sentence_transformers.py

@@ -52,6 +52,9 @@ async def shutdown(self) -> None:
     async def should_refresh_models(self) -> bool:
         return False
 
+    async def allow_listing_models(self) -> bool:
+        return True
+
     async def list_models(self) -> list[Model] | None:
         return [
             Model(

llama_stack/providers/remote/inference/vllm/config.py

@@ -34,6 +34,10 @@ class VLLMInferenceAdapterConfig(RemoteInferenceProviderConfig):
         default=False,
         description="Whether to refresh models periodically",
     )
+    allow_listing_models: bool = Field(
+        default=True,
+        description="Whether to allow listing models from the vLLM server",
+    )
 
     @field_validator("tls_verify")
     @classmethod
@@ -59,4 +63,5 @@ def sample_run_config(
             "max_tokens": "${env.VLLM_MAX_TOKENS:=4096}",
             "api_token": "${env.VLLM_API_TOKEN:=fake}",
             "tls_verify": "${env.VLLM_TLS_VERIFY:=true}",
+            "allow_listing_models": "${env.VLLM_ALLOW_LISTING_MODELS:=true}",
         }

llama_stack/providers/remote/inference/vllm/vllm.py

@@ -282,7 +282,18 @@ async def should_refresh_models(self) -> bool:
         # Strictly respecting the refresh_models directive
         return self.config.refresh_models
 
+    async def allow_listing_models(self) -> bool:
+        # Respecting the allow_listing_models directive
+        result = self.config.allow_listing_models
+        log.debug(f"VLLM allow_listing_models: {result}")
+        return result
+
     async def list_models(self) -> list[Model] | None:
+        log.debug(f"VLLM list_models called, allow_listing_models={self.config.allow_listing_models}")
+        if not self.config.allow_listing_models:
+            log.debug("VLLM list_models returning None due to allow_listing_models=False")
+            return None
+
         models = []
         async for m in self.client.models.list():
             model_type = ModelType.llm  # unclear how to determine embedding vs. llm models
@@ -332,24 +343,34 @@ async def _get_model(self, model_id: str) -> Model:
     def get_extra_client_params(self):
         return {"http_client": httpx.AsyncClient(verify=self.config.tls_verify)}
 
-    async def register_model(self, model: Model) -> Model:
-        try:
-            model = await self.register_helper.register_model(model)
-        except ValueError:
-            pass  # Ignore statically unknown model, will check live listing
+    async def check_model_availability(self, model: str) -> bool:
+        """
+        Check if a specific model is available from the vLLM server.
+
+        This method respects the allow_listing_models configuration flag.
+        If allow_listing_models is False, it returns True to allow model registration
+        without making HTTP requests (trusting that the model exists).
+
+        :param model: The model identifier to check.
+        :return: True if the model is available or if allow_listing_models is False, False otherwise.
+        """
+        # Check if provider allows listing models before making HTTP request
+        if not self.config.allow_listing_models:
+            log.debug(
+                "VLLM check_model_availability returning True due to allow_listing_models=False (trusting model exists)"
+            )
+            return True
+
         try:
             res = self.client.models.list()
         except APIConnectionError as e:
-            raise ValueError(
-                f"Failed to connect to vLLM at {self.config.url}. Please check if vLLM is running and accessible at that URL."
-            ) from e
+            log.warning(f"Failed to connect to vLLM at {self.config.url}: {e}")
+            return False
+
         available_models = [m.id async for m in res]
-        if model.provider_resource_id not in available_models:
-            raise ValueError(
-                f"Model {model.provider_resource_id} is not being served by vLLM. "
-                f"Available models: {', '.join(available_models)}"
-            )
-        return model
+        is_available = model in available_models
+        log.debug(f"VLLM model {model} availability: {is_available}")
+        return is_available
 
     async def _get_params(self, request: ChatCompletionRequest) -> dict:
         options = get_sampling_options(request.sampling_params)

llama_stack/providers/utils/inference/model_registry.py

@@ -100,6 +100,9 @@ async def list_models(self) -> list[Model] | None:
     async def should_refresh_models(self) -> bool:
         return False
 
+    async def allow_listing_models(self) -> bool:
+        return True
+
     def get_provider_model_id(self, identifier: str) -> str | None:
         return self.alias_to_provider_id_map.get(identifier, None)
 

llama_stack/providers/utils/inference/openai_mixin.py

@@ -425,3 +425,6 @@ async def check_model_availability(self, model: str) -> bool:
 
     async def should_refresh_models(self) -> bool:
         return False
+
+    async def allow_listing_models(self) -> bool:
+        return True

tests/unit/distribution/routers/test_routing_tables.py

@@ -52,6 +52,9 @@ async def unregister_model(self, model_id: str):
     async def should_refresh_models(self):
         return False
 
+    async def allow_listing_models(self):
+        return True
+
     async def list_models(self):
         return [
             Model(

tests/unit/providers/inference/test_remote_vllm.py

@@ -636,27 +636,75 @@ async def test_should_refresh_models():
     Test the should_refresh_models method with different refresh_models configurations.
 
     This test verifies that:
-    1. When refresh_models is True, should_refresh_models returns True regardless of api_token
-    2. When refresh_models is False, should_refresh_models returns False regardless of api_token
+    1. When refresh_models is True, should_refresh_models returns True
+    2. When refresh_models is False, should_refresh_models returns False
     """
 
-    # Test case 1: refresh_models is True, api_token is None
-    config1 = VLLMInferenceAdapterConfig(url="http://test.localhost", api_token=None, refresh_models=True)
+    # Test case 1: refresh_models is True
+    config1 = VLLMInferenceAdapterConfig(url="http://test.localhost", refresh_models=True)
     adapter1 = VLLMInferenceAdapter(config1)
     result1 = await adapter1.should_refresh_models()
     assert result1 is True, "should_refresh_models should return True when refresh_models is True"
 
-    # Test case 2: refresh_models is True, api_token is empty string
-    config2 = VLLMInferenceAdapterConfig(url="http://test.localhost", api_token="", refresh_models=True)
+    # Test case 2: refresh_models is False
+    config2 = VLLMInferenceAdapterConfig(url="http://test.localhost", refresh_models=False)
     adapter2 = VLLMInferenceAdapter(config2)
     result2 = await adapter2.should_refresh_models()
-    assert result2 is True, "should_refresh_models should return True when refresh_models is True"
+    assert result2 is False, "should_refresh_models should return False when refresh_models is False"
 
-    # Test case 3: refresh_models is True, api_token is "fake" (default)
-    config3 = VLLMInferenceAdapterConfig(url="http://test.localhost", api_token="fake", refresh_models=True)
+
+async def test_allow_listing_models_flag():
+    """
+    Test the allow_listing_models flag functionality.
+
+    This test verifies that:
+    1. When allow_listing_models is True (default), list_models returns models from the server
+    2. When allow_listing_models is False, list_models returns None without calling the server
+    """
+
+    # Test case 1: allow_listing_models is True (default)
+    config1 = VLLMInferenceAdapterConfig(url="http://test.localhost", allow_listing_models=True)
+    adapter1 = VLLMInferenceAdapter(config1)
+    adapter1.__provider_id__ = "test-vllm"
+
+    # Mock the client.models.list() method
+    with patch.object(VLLMInferenceAdapter, "client", new_callable=PropertyMock) as mock_client_property:
+        mock_client = MagicMock()
+
+        async def mock_models_list():
+            yield OpenAIModel(id="test-model-1", created=1, object="model", owned_by="test")
+            yield OpenAIModel(id="test-model-2", created=2, object="model", owned_by="test")
+
+        mock_client.models.list.return_value = mock_models_list()
+        mock_client_property.return_value = mock_client
+
+        models = await adapter1.list_models()
+        assert models is not None, "list_models should return models when allow_listing_models is True"
+        assert len(models) == 2, "Should return 2 models"
+        assert models[0].identifier == "test-model-1"
+        assert models[1].identifier == "test-model-2"
+        mock_client.models.list.assert_called_once()
+
+    # Test case 2: allow_listing_models is False
+    config2 = VLLMInferenceAdapterConfig(url="http://test.localhost", allow_listing_models=False)
+    adapter2 = VLLMInferenceAdapter(config2)
+    adapter2.__provider_id__ = "test-vllm"
+
+    # Mock the client.models.list() method
+    with patch.object(VLLMInferenceAdapter, "client", new_callable=PropertyMock) as mock_client_property:
+        mock_client = MagicMock()
+        mock_client_property.return_value = mock_client
+
+        models = await adapter2.list_models()
+        assert models is None, "list_models should return None when allow_listing_models is False"
+        mock_client.models.list.assert_not_called()
+
+    # Test case 3: allow_listing_models defaults to True
+    config3 = VLLMInferenceAdapterConfig(url="http://test.localhost")
     adapter3 = VLLMInferenceAdapter(config3)
-    result3 = await adapter3.should_refresh_models()
-    assert result3 is True, "should_refresh_models should return True when refresh_models is True"
+    adapter3.__provider_id__ = "test-vllm"
+    result3 = await adapter3.allow_listing_models()
+    assert result3 is True, "allow_listing_models should return True by default"
 
     # Test case 4: refresh_models is True, api_token is real token
     config4 = VLLMInferenceAdapterConfig(url="http://test.localhost", api_token="real-token-123", refresh_models=True)
@@ -670,6 +718,22 @@ async def test_should_refresh_models():
     result5 = await adapter5.should_refresh_models()
     assert result5 is False, "should_refresh_models should return False when refresh_models is False"
 
+    # Mock the client.models.list() method
+    with patch.object(VLLMInferenceAdapter, "client", new_callable=PropertyMock) as mock_client_property:
+        mock_client = MagicMock()
+
+        async def mock_models_list():
+            yield OpenAIModel(id="default-model", created=1, object="model", owned_by="test")
+
+        mock_client.models.list.return_value = mock_models_list()
+        mock_client_property.return_value = mock_client
+
+        models = await adapter3.list_models()
+        assert models is not None, "list_models should return models when allow_listing_models defaults to True"
+        assert len(models) == 1, "Should return 1 model"
+        assert models[0].identifier == "default-model"
+        mock_client.models.list.assert_called_once()
+
 
 async def test_provider_data_var_context_propagation(vllm_inference_adapter):
     """

tests/unit/server/test_access_control.py

@@ -32,6 +32,7 @@ async def test_setup(cached_disk_dist_registry):
     mock_inference.__provider_spec__ = MagicMock()
     mock_inference.__provider_spec__.api = Api.inference
     mock_inference.register_model = AsyncMock(side_effect=_return_model)
+    mock_inference.allow_listing_models = AsyncMock(return_value=True)
     routing_table = ModelsRoutingTable(
         impls_by_provider_id={"test_provider": mock_inference},
         dist_registry=cached_disk_dist_registry,