feat(vllm): periodically refresh models (#2823)

Just like #2805 but for vLLM.

We also make VLLM_URL env variable optional (not required) -- if not
specified, the provider silently sits idle and yells eventually if
someone tries to call a completion on it. This is done so as to allow
this provider to be present in the `starter` distribution.

## Test Plan

Set up vLLM, copy the starter template and set `{ refresh_models: true,
refresh_models_interval: 10 }` for the vllm provider and then run:

```
ENABLE_VLLM=vllm VLLM_URL=http://localhost:8000/v1 \
  uv run llama stack run --image-type venv /tmp/starter.yaml
```

Verify that `llama-stack-client models list` brings up the model
correctly from vLLM.

docs/source/providers/inference/remote_vllm.md

@@ -12,11 +12,13 @@ Remote vLLM inference provider for connecting to vLLM servers.
 | `max_tokens` | `<class 'int'>` | No | 4096 | Maximum number of tokens to generate. |
 | `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 |
+| `refresh_models_interval` | `<class 'int'>` | No | 300 | Interval in seconds to refresh models |
 
 ## Sample Configuration
 
 ```yaml
-url: ${env.VLLM_URL}
+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}

llama_stack/apis/inference/inference.py

@@ -819,7 +819,7 @@ class OpenAIEmbeddingsResponse(BaseModel):
 class ModelStore(Protocol):
     async def get_model(self, identifier: str) -> Model: ...
 
-    async def update_registered_models(
+    async def update_registered_llm_models(
         self,
         provider_id: str,
         models: list[Model],

llama_stack/distribution/routing_tables/models.py

@@ -81,7 +81,7 @@ class ModelsRoutingTable(CommonRoutingTableImpl, Models):
             raise ValueError(f"Model {model_id} not found")
         await self.unregister_object(existing_model)
 
-    async def update_registered_models(
+    async def update_registered_llm_models(
         self,
         provider_id: str,
         models: list[Model],
@@ -92,12 +92,16 @@ class ModelsRoutingTable(CommonRoutingTableImpl, Models):
         # from run.yaml) that we need to keep track of
         model_ids = {}
         for model in existing_models:
-            if model.provider_id == provider_id:
+            # we leave embeddings models alone because often we don't get metadata
+            # (embedding dimension, etc.) from the provider
+            if model.provider_id == provider_id and model.model_type == ModelType.llm:
                 model_ids[model.provider_resource_id] = model.identifier
                 logger.debug(f"unregistering model {model.identifier}")
                 await self.unregister_object(model)
 
         for model in models:
+            if model.model_type != ModelType.llm:
+                continue
             if model.provider_resource_id in model_ids:
                 model.identifier = model_ids[model.provider_resource_id]
 

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

@@ -159,18 +159,18 @@ class OllamaInferenceAdapter(
             models = []
             for m in response.models:
                 model_type = ModelType.embedding if m.details.family in ["bert"] else ModelType.llm
-                # unfortunately, ollama does not provide embedding dimension in the model list :(
+                if model_type == ModelType.embedding:
-                # we should likely add a hard-coded mapping of model name to embedding dimension
+                    continue
                 models.append(
                     Model(
                         identifier=m.model,
                         provider_resource_id=m.model,
                         provider_id=provider_id,
-                        metadata={"embedding_dimension": 384} if model_type == ModelType.embedding else {},
+                        metadata={},
                         model_type=model_type,
                     )
                 )
-            await self.model_store.update_registered_models(provider_id, models)
+            await self.model_store.update_registered_llm_models(provider_id, models)
             logger.debug(f"ollama refreshed model list ({len(models)} models)")
 
             await asyncio.sleep(self.config.refresh_models_interval)

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

@@ -29,6 +29,14 @@ class VLLMInferenceAdapterConfig(BaseModel):
         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",
+    )
+    refresh_models_interval: int = Field(
+        default=300,
+        description="Interval in seconds to refresh models",
+    )
 
     @field_validator("tls_verify")
     @classmethod
@@ -46,7 +54,7 @@ class VLLMInferenceAdapterConfig(BaseModel):
     @classmethod
     def sample_run_config(
         cls,
-        url: str = "${env.VLLM_URL}",
+        url: str = "${env.VLLM_URL:=}",
         **kwargs,
     ):
         return {

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

@@ -3,8 +3,8 @@
 #
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
+import asyncio
 import json
-import logging
 from collections.abc import AsyncGenerator, AsyncIterator
 from typing import Any
 
@@ -38,6 +38,7 @@ from llama_stack.apis.inference import (
     JsonSchemaResponseFormat,
     LogProbConfig,
     Message,
+    ModelStore,
     OpenAIChatCompletion,
     OpenAICompletion,
     OpenAIEmbeddingData,
@@ -54,6 +55,7 @@ from llama_stack.apis.inference import (
     ToolPromptFormat,
 )
 from llama_stack.apis.models import Model, ModelType
+from llama_stack.log import get_logger
 from llama_stack.models.llama.datatypes import BuiltinTool, StopReason, ToolCall
 from llama_stack.models.llama.sku_list import all_registered_models
 from llama_stack.providers.datatypes import (
@@ -84,7 +86,7 @@ from llama_stack.providers.utils.inference.prompt_adapter import (
 
 from .config import VLLMInferenceAdapterConfig
 
-log = logging.getLogger(__name__)
+log = get_logger(name=__name__, category="inference")
 
 
 def build_hf_repo_model_entries():
@@ -288,16 +290,76 @@ async def _process_vllm_chat_completion_stream_response(
 
 
 class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
+    # automatically set by the resolver when instantiating the provider
+    __provider_id__: str
+    model_store: ModelStore | None = None
+    _refresh_task: asyncio.Task | None = None
+
     def __init__(self, config: VLLMInferenceAdapterConfig) -> None:
         self.register_helper = ModelRegistryHelper(build_hf_repo_model_entries())
         self.config = config
         self.client = None
 
     async def initialize(self) -> None:
-        pass
+        if not self.config.url:
+            # intentionally don't raise an error here, we want to allow the provider to be "dormant"
+            # or available in distributions like "starter" without causing a ruckus
+            return
+
+        if self.config.refresh_models:
+            self._refresh_task = asyncio.create_task(self._refresh_models())
+
+            def cb(task):
+                import traceback
+
+                if task.cancelled():
+                    log.error(f"vLLM background refresh task canceled:\n{''.join(traceback.format_stack())}")
+                elif task.exception():
+                    # print the stack trace for the exception
+                    exc = task.exception()
+                    log.error(f"vLLM background refresh task died: {exc}")
+                    traceback.print_exception(exc)
+                else:
+                    log.error("vLLM background refresh task completed unexpectedly")
+
+            self._refresh_task.add_done_callback(cb)
+
+    async def _refresh_models(self) -> None:
+        provider_id = self.__provider_id__
+        waited_time = 0
+        while not self.model_store and waited_time < 60:
+            await asyncio.sleep(1)
+            waited_time += 1
+
+        if not self.model_store:
+            raise ValueError("Model store not set after waiting 60 seconds")
+
+        self._lazy_initialize_client()
+        assert self.client is not None  # mypy
+        while True:
+            try:
+                models = []
+                async for m in self.client.models.list():
+                    model_type = ModelType.llm  # unclear how to determine embedding vs. llm models
+                    models.append(
+                        Model(
+                            identifier=m.id,
+                            provider_resource_id=m.id,
+                            provider_id=provider_id,
+                            metadata={},
+                            model_type=model_type,
+                        )
+                    )
+                await self.model_store.update_registered_llm_models(provider_id, models)
+                log.debug(f"vLLM refreshed model list ({len(models)} models)")
+            except Exception as e:
+                log.error(f"vLLM background refresh task failed: {e}")
+            await asyncio.sleep(self.config.refresh_models_interval)
 
     async def shutdown(self) -> None:
-        pass
+        if self._refresh_task:
+            self._refresh_task.cancel()
+            self._refresh_task = None
 
     async def unregister_model(self, model_id: str) -> None:
         pass
@@ -312,6 +374,9 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
             HealthResponse: A dictionary containing the health status.
         """
         try:
+            if not self.config.url:
+                return HealthResponse(status=HealthStatus.ERROR, message="vLLM URL is not set")
+
             client = self._create_client() if self.client is None else self.client
             _ = [m async for m in client.models.list()]  # Ensure the client is initialized
             return HealthResponse(status=HealthStatus.OK)
@@ -327,6 +392,11 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
         if self.client is not None:
             return
 
+        if not self.config.url:
+            raise ValueError(
+                "You must provide a vLLM URL in the run.yaml file (or set the VLLM_URL environment variable)"
+            )
+
         log.info(f"Initializing vLLM client with base_url={self.config.url}")
         self.client = self._create_client()
 

llama_stack/templates/starter/run.yaml

@@ -26,7 +26,7 @@ providers:
   - provider_id: ${env.ENABLE_VLLM:=__disabled__}
     provider_type: remote::vllm
     config:
-      url: ${env.VLLM_URL}
+      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}