feat: expand set of known openai models, allow using openai canonical model names (#2164)

note: the openai provider exposes the litellm specific model names to
the user. this change is compatible with that. the litellm names should
be deprecated.

llama_stack/providers/remote/inference/openai/models.py

@@ -4,27 +4,60 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from dataclasses import dataclass
+
 from llama_stack.apis.models.models import ModelType
 from llama_stack.providers.utils.inference.model_registry import (
     ProviderModelEntry,
 )
 
 LLM_MODEL_IDS = [
+    # the models w/ "openai/" prefix are the litellm specific model names.
+    # they should be deprecated in favor of the canonical openai model names.
     "openai/gpt-4o",
     "openai/gpt-4o-mini",
     "openai/chatgpt-4o-latest",
+    "gpt-3.5-turbo-0125",
+    "gpt-3.5-turbo",
+    "gpt-3.5-turbo-instruct",
+    "gpt-4",
+    "gpt-4-turbo",
+    "gpt-4o",
+    "gpt-4o-2024-08-06",
+    "gpt-4o-mini",
+    "gpt-4o-audio-preview",
+    "chatgpt-4o-latest",
+    "o1",
+    "o1-mini",
+    "o3-mini",
+    "o4-mini",
 ]
 
 
+@dataclass
+class EmbeddingModelInfo:
+    """Structured representation of embedding model information."""
+
+    embedding_dimension: int
+    context_length: int
+
+
+EMBEDDING_MODEL_IDS: dict[str, EmbeddingModelInfo] = {
+    "openai/text-embedding-3-small": EmbeddingModelInfo(1536, 8192),
+    "openai/text-embedding-3-large": EmbeddingModelInfo(3072, 8192),
+    "text-embedding-3-small": EmbeddingModelInfo(1536, 8192),
+    "text-embedding-3-large": EmbeddingModelInfo(3072, 8192),
+}
+
+
 MODEL_ENTRIES = [ProviderModelEntry(provider_model_id=m) for m in LLM_MODEL_IDS] + [
     ProviderModelEntry(
-        provider_model_id="openai/text-embedding-3-small",
+        provider_model_id=model_id,
         model_type=ModelType.embedding,
-        metadata={"embedding_dimension": 1536, "context_length": 8192},
-    ),
-    ProviderModelEntry(
-        provider_model_id="openai/text-embedding-3-large",
-        model_type=ModelType.embedding,
-        metadata={"embedding_dimension": 3072, "context_length": 8192},
-    ),
+        metadata={
+            "embedding_dimension": model_info.embedding_dimension,
+            "context_length": model_info.context_length,
+        },
+    )
+    for model_id, model_info in EMBEDDING_MODEL_IDS.items()
 ]

llama_stack/providers/remote/inference/openai/openai.py

@@ -19,6 +19,13 @@ class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
             provider_data_api_key_field="openai_api_key",
         )
         self.config = config
+        # we set is_openai_compat so users can use the canonical
+        # openai model names like "gpt-4" or "gpt-3.5-turbo"
+        # and the model name will be translated to litellm's
+        # "openai/gpt-4" or "openai/gpt-3.5-turbo" transparently.
+        # if we do not set this, users will be exposed to the
+        # litellm specific model names, an abstraction leak.
+        self.is_openai_compat = True
 
     async def initialize(self) -> None:
         await super().initialize()

llama_stack/providers/utils/inference/litellm_openai_mixin.py

@@ -62,6 +62,9 @@ class LiteLLMOpenAIMixin(
     Inference,
     NeedsRequestProviderData,
 ):
+    # TODO: avoid exposing the litellm specific model names to the user.
+    #       potential change: add a prefix param that gets added to the model name
+    #                         when calling litellm.
     def __init__(
         self,
         model_entries,
@@ -92,7 +95,9 @@ class LiteLLMOpenAIMixin(
         return model
 
     def get_litellm_model_name(self, model_id: str) -> str:
-        return "openai/" + model_id if self.is_openai_compat else model_id
+        # 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
 
     async def completion(
         self,