feat(starter)!: simplify starter distro; litellm model registry changes (#2916)

llama_stack/providers/utils/inference/litellm_openai_mixin.py

@@ -68,11 +68,14 @@ class LiteLLMOpenAIMixin(
     def __init__(
         self,
         model_entries,
+        litellm_provider_name: str,
         api_key_from_config: str | None,
         provider_data_api_key_field: str,
         openai_compat_api_base: str | None = None,
     ):
         ModelRegistryHelper.__init__(self, model_entries)
+
+        self.litellm_provider_name = litellm_provider_name
         self.api_key_from_config = api_key_from_config
         self.provider_data_api_key_field = provider_data_api_key_field
         self.api_base = openai_compat_api_base
@@ -91,7 +94,11 @@ class LiteLLMOpenAIMixin(
     def get_litellm_model_name(self, model_id: str) -> str:
         # users may be using openai/ prefix in their model names. the openai/models.py did this by default.
         # model_id.startswith("openai/") is for backwards compatibility.
-        return "openai/" + model_id if self.is_openai_compat and not model_id.startswith("openai/") else model_id
+        return (
+            f"{self.litellm_provider_name}/{model_id}"
+            if self.is_openai_compat and not model_id.startswith(self.litellm_provider_name)
+            else model_id
+        )
 
     async def completion(
         self,

llama_stack/providers/utils/inference/model_registry.py

@@ -50,7 +50,8 @@ def build_hf_repo_model_entry(
     additional_aliases: list[str] | None = None,
 ) -> ProviderModelEntry:
     aliases = [
-        get_huggingface_repo(model_descriptor),
+        # NOTE: avoid HF aliases because they _cannot_ be unique across providers
+        # get_huggingface_repo(model_descriptor),
     ]
     if additional_aliases:
         aliases.extend(additional_aliases)
@@ -75,7 +76,9 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
     __provider_id__: str
 
     def __init__(self, model_entries: list[ProviderModelEntry], allowed_models: list[str] | None = None):
+        self.model_entries = model_entries
         self.allowed_models = allowed_models
+
         self.alias_to_provider_id_map = {}
         self.provider_id_to_llama_model_map = {}
         for entry in model_entries:
@@ -98,7 +101,7 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
                     continue
                 models.append(
                     Model(
-                        model_id=id,
+                        identifier=id,
                         provider_resource_id=entry.provider_model_id,
                         model_type=ModelType.llm,
                         metadata=entry.metadata,
@@ -185,8 +188,8 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
         return model
 
     async def unregister_model(self, model_id: str) -> None:
-        # TODO: should we block unregistering base supported provider model IDs?
-        if model_id not in self.alias_to_provider_id_map:
-            raise ValueError(f"Model id '{model_id}' is not registered.")
-
-        del self.alias_to_provider_id_map[model_id]
+        # model_id is the identifier, not the provider_resource_id
+        # unfortunately, this ID can be of the form provider_id/model_id which
+        # we never registered. TODO: fix this by significantly rewriting
+        # registration and registry helper
+        pass