feat: add refresh_models support to inference adapters (default: false) (#3719)

# What does this PR do?

inference adapters can now configure `refresh_models: bool` to control
periodic model listing from their providers

BREAKING CHANGE: together inference adapter default changed. previously
always refreshed, now follows config.

addresses "models: refresh" on #3517

## Test Plan

ci w/ new tests

llama_stack/providers/remote/inference/databricks/databricks.py

@@ -41,9 +41,6 @@ class DatabricksInferenceAdapter(OpenAIMixin):
             ).serving_endpoints.list()  # TODO: this is not async
         ]
 
-    async def should_refresh_models(self) -> bool:
-        return False
-
     async def openai_completion(
         self,
         model: str,

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

@@ -6,8 +6,6 @@
 
 from typing import Any
 
-from pydantic import Field
-
 from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
 
 DEFAULT_OLLAMA_URL = "http://localhost:11434"
@@ -15,10 +13,6 @@ DEFAULT_OLLAMA_URL = "http://localhost:11434"
 
 class OllamaImplConfig(RemoteInferenceProviderConfig):
     url: str = DEFAULT_OLLAMA_URL
-    refresh_models: bool = Field(
-        default=False,
-        description="Whether to refresh models periodically",
-    )
 
     @classmethod
     def sample_run_config(cls, url: str = "${env.OLLAMA_URL:=http://localhost:11434}", **kwargs) -> dict[str, Any]:

llama_stack/providers/remote/inference/ollama/ollama.py

@@ -72,9 +72,6 @@ class OllamaInferenceAdapter(OpenAIMixin):
                 f"Ollama Server is not running (message: {r['message']}). Make sure to start it using `ollama serve` in a separate terminal"
             )
 
-    async def should_refresh_models(self) -> bool:
-        return self.config.refresh_models
-
     async def health(self) -> HealthResponse:
         """
         Performs a health check by verifying connectivity to the Ollama server.

llama_stack/providers/remote/inference/together/together.py

@@ -63,9 +63,6 @@ class TogetherInferenceAdapter(OpenAIMixin, NeedsRequestProviderData):
         # Together's /v1/models is not compatible with OpenAI's /v1/models. Together support ticket #13355 -> will not fix, use Together's own client
         return [m.id for m in await self._get_client().models.list()]
 
-    async def should_refresh_models(self) -> bool:
-        return True
-
     async def openai_embeddings(
         self,
         model: str,

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

@@ -30,10 +30,6 @@ class VLLMInferenceAdapterConfig(RemoteInferenceProviderConfig):
         default=True,
         description="Whether to verify TLS certificates. Can be a boolean or a path to a CA certificate file.",
     )
-    refresh_models: bool = Field(
-        default=False,
-        description="Whether to refresh models periodically",
-    )
 
     @field_validator("tls_verify")
     @classmethod

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

@@ -53,10 +53,6 @@ class VLLMInferenceAdapter(OpenAIMixin):
                 "You must provide a URL in run.yaml (or via the VLLM_URL environment variable) to use vLLM."
             )
 
-    async def should_refresh_models(self) -> bool:
-        # Strictly respecting the refresh_models directive
-        return self.config.refresh_models
-
     async def health(self) -> HealthResponse:
         """
         Performs a health check by verifying connectivity to the remote vLLM server.

llama_stack/providers/utils/inference/model_registry.py

@@ -24,6 +24,10 @@ class RemoteInferenceProviderConfig(BaseModel):
         default=None,
         description="List of models that should be registered with the model registry. If None, all models are allowed.",
     )
+    refresh_models: bool = Field(
+        default=False,
+        description="Whether to refresh models periodically from the provider",
+    )
 
 
 # TODO: this class is more confusing than useful right now. We need to make it

llama_stack/providers/utils/inference/openai_mixin.py

@@ -484,7 +484,7 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
         return model in self._model_cache
 
     async def should_refresh_models(self) -> bool:
-        return False
+        return self.config.refresh_models
 
     #
     # The model_dump implementations are to avoid serializing the extra fields,