feat: add (openai, anthropic, gemini) providers via litellm (#1267)

# What does this PR do?

This PR introduces more non-llama model support to llama stack.
Providers introduced: openai, anthropic and gemini. All of these
providers use essentially the same piece of code -- the implementation
works via the `litellm` library.

We will expose only specific models for providers we enable making sure
they all work well and pass tests. This setup (instead of automatically
enabling _all_ providers and models allowed by LiteLLM) ensures we can
also perform any needed prompt tuning on a per-model basis as needed
(just like we do it for llama models.)

## Test Plan

```bash
#!/bin/bash

args=("$@")
for model in openai/gpt-4o anthropic/claude-3-5-sonnet-latest gemini/gemini-1.5-flash; do
    LLAMA_STACK_CONFIG=dev pytest -s -v tests/client-sdk/inference/test_text_inference.py \
        --embedding-model=all-MiniLM-L6-v2 \
        --vision-inference-model="" \
        --inference-model=$model "${args[@]}"
done
```

llama_stack/providers/registry/inference.py

@@ -207,6 +207,33 @@ def available_providers() -> List[ProviderSpec]:
                 config_class="llama_stack.providers.remote.inference.runpod.RunpodImplConfig",
             ),
         ),
+        remote_provider_spec(
+            api=Api.inference,
+            adapter=AdapterSpec(
+                adapter_type="openai",
+                pip_packages=["litellm"],
+                module="llama_stack.providers.remote.inference.openai",
+                config_class="llama_stack.providers.remote.inference.openai.OpenAIConfig",
+            ),
+        ),
+        remote_provider_spec(
+            api=Api.inference,
+            adapter=AdapterSpec(
+                adapter_type="anthropic",
+                pip_packages=["litellm"],
+                module="llama_stack.providers.remote.inference.anthropic",
+                config_class="llama_stack.providers.remote.inference.anthropic.AnthropicConfig",
+            ),
+        ),
+        remote_provider_spec(
+            api=Api.inference,
+            adapter=AdapterSpec(
+                adapter_type="gemini",
+                pip_packages=["litellm"],
+                module="llama_stack.providers.remote.inference.gemini",
+                config_class="llama_stack.providers.remote.inference.gemini.GeminiConfig",
+            ),
+        ),
         remote_provider_spec(
             api=Api.inference,
             adapter=AdapterSpec(

llama_stack/providers/remote/inference/anthropic/__init__.py

@@ -0,0 +1,23 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Optional
+
+from pydantic import BaseModel
+
+from .config import AnthropicConfig
+
+
+class AnthropicProviderDataValidator(BaseModel):
+    anthropic_api_key: Optional[str] = None
+
+
+async def get_adapter_impl(config: AnthropicConfig, _deps):
+    from .anthropic import AnthropicInferenceAdapter
+
+    impl = AnthropicInferenceAdapter(config)
+    await impl.initialize()
+    return impl

llama_stack/providers/remote/inference/anthropic/anthropic.py

@@ -0,0 +1,22 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from llama_stack.providers.utils.inference.litellm_openai_mixin import LiteLLMOpenAIMixin
+
+from .config import AnthropicConfig
+from .models import MODEL_ENTRIES
+
+
+class AnthropicInferenceAdapter(LiteLLMOpenAIMixin):
+    def __init__(self, config: AnthropicConfig) -> None:
+        LiteLLMOpenAIMixin.__init__(self, MODEL_ENTRIES)
+        self.config = config
+
+    async def initialize(self) -> None:
+        pass
+
+    async def shutdown(self) -> None:
+        pass

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

@@ -0,0 +1,25 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any, Dict, Optional
+
+from pydantic import BaseModel, Field
+
+from llama_stack.schema_utils import json_schema_type
+
+
+@json_schema_type
+class AnthropicConfig(BaseModel):
+    api_key: Optional[str] = Field(
+        default=None,
+        description="API key for Anthropic models",
+    )
+
+    @classmethod
+    def sample_run_config(cls, api_key: str = "${env.ANTHROPIC_API_KEY}", **kwargs) -> Dict[str, Any]:
+        return {
+            "api_key": api_key,
+        }

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

@@ -0,0 +1,35 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from llama_stack.apis.models.models import ModelType
+from llama_stack.providers.utils.inference.model_registry import (
+    ProviderModelEntry,
+)
+
+LLM_MODEL_IDS = [
+    "anthropic/claude-3-5-sonnet-latest",
+    "anthropic/claude-3-7-sonnet-latest",
+    "anthropic/claude-3-5-haiku-latest",
+]
+
+
+MODEL_ENTRIES = [ProviderModelEntry(provider_model_id=m) for m in LLM_MODEL_IDS] + [
+    ProviderModelEntry(
+        provider_model_id="anthropic/voyage-3",
+        model_type=ModelType.embedding,
+        metadata={"embedding_dimension": 1024, "context_length": 32000},
+    ),
+    ProviderModelEntry(
+        provider_model_id="anthropic/voyage-3-lite",
+        model_type=ModelType.embedding,
+        metadata={"embedding_dimension": 512, "context_length": 32000},
+    ),
+    ProviderModelEntry(
+        provider_model_id="anthropic/voyage-code-3",
+        model_type=ModelType.embedding,
+        metadata={"embedding_dimension": 1024, "context_length": 32000},
+    ),
+]

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

@@ -23,8 +23,8 @@ class FireworksImplConfig(BaseModel):
     )
 
     @classmethod
-    def sample_run_config(cls, **kwargs) -> Dict[str, Any]:
+    def sample_run_config(cls, api_key: str = "${env.FIREWORKS_API_KEY}", **kwargs) -> Dict[str, Any]:
         return {
             "url": "https://api.fireworks.ai/inference/v1",
-            "api_key": "${env.FIREWORKS_API_KEY}",
+            "api_key": api_key,
         }

llama_stack/providers/remote/inference/gemini/__init__.py

@@ -0,0 +1,23 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Optional
+
+from pydantic import BaseModel
+
+from .config import GeminiConfig
+
+
+class GeminiProviderDataValidator(BaseModel):
+    gemini_api_key: Optional[str] = None
+
+
+async def get_adapter_impl(config: GeminiConfig, _deps):
+    from .gemini import GeminiInferenceAdapter
+
+    impl = GeminiInferenceAdapter(config)
+    await impl.initialize()
+    return impl

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

@@ -0,0 +1,25 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any, Dict, Optional
+
+from pydantic import BaseModel, Field
+
+from llama_stack.schema_utils import json_schema_type
+
+
+@json_schema_type
+class GeminiConfig(BaseModel):
+    api_key: Optional[str] = Field(
+        default=None,
+        description="API key for Gemini models",
+    )
+
+    @classmethod
+    def sample_run_config(cls, api_key: str = "${env.GEMINI_API_KEY}", **kwargs) -> Dict[str, Any]:
+        return {
+            "api_key": api_key,
+        }

llama_stack/providers/remote/inference/gemini/gemini.py

@@ -0,0 +1,22 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from llama_stack.providers.utils.inference.litellm_openai_mixin import LiteLLMOpenAIMixin
+
+from .config import GeminiConfig
+from .models import MODEL_ENTRIES
+
+
+class GeminiInferenceAdapter(LiteLLMOpenAIMixin):
+    def __init__(self, config: GeminiConfig) -> None:
+        LiteLLMOpenAIMixin.__init__(self, MODEL_ENTRIES)
+        self.config = config
+
+    async def initialize(self) -> None:
+        pass
+
+    async def shutdown(self) -> None:
+        pass

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

@@ -0,0 +1,24 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from llama_stack.apis.models.models import ModelType
+from llama_stack.providers.utils.inference.model_registry import (
+    ProviderModelEntry,
+)
+
+LLM_MODEL_IDS = [
+    "gemini/gemini-1.5-flash",
+    "gemini/gemini-1.5-pro",
+]
+
+
+MODEL_ENTRIES = [ProviderModelEntry(provider_model_id=m) for m in LLM_MODEL_IDS] + [
+    ProviderModelEntry(
+        provider_model_id="gemini/text-embedding-004",
+        model_type=ModelType.embedding,
+        metadata={"embedding_dimension": 768, "context_length": 2048},
+    ),
+]

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

@@ -0,0 +1,23 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Optional
+
+from pydantic import BaseModel
+
+from .config import OpenAIConfig
+
+
+class OpenAIProviderDataValidator(BaseModel):
+    openai_api_key: Optional[str] = None
+
+
+async def get_adapter_impl(config: OpenAIConfig, _deps):
+    from .openai import OpenAIInferenceAdapter
+
+    impl = OpenAIInferenceAdapter(config)
+    await impl.initialize()
+    return impl

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

@@ -0,0 +1,25 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import Any, Dict, Optional
+
+from pydantic import BaseModel, Field
+
+from llama_stack.schema_utils import json_schema_type
+
+
+@json_schema_type
+class OpenAIConfig(BaseModel):
+    api_key: Optional[str] = Field(
+        default=None,
+        description="API key for OpenAI models",
+    )
+
+    @classmethod
+    def sample_run_config(cls, api_key: str = "${env.OPENAI_API_KEY}", **kwargs) -> Dict[str, Any]:
+        return {
+            "api_key": api_key,
+        }

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

@@ -0,0 +1,30 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from llama_stack.apis.models.models import ModelType
+from llama_stack.providers.utils.inference.model_registry import (
+    ProviderModelEntry,
+)
+
+LLM_MODEL_IDS = [
+    "openai/gpt-4o",
+    "openai/gpt-4o-mini",
+    "openai/chatgpt-4o-latest",
+]
+
+
+MODEL_ENTRIES = [ProviderModelEntry(provider_model_id=m) for m in LLM_MODEL_IDS] + [
+    ProviderModelEntry(
+        provider_model_id="openai/text-embedding-3-small",
+        model_type=ModelType.embedding,
+        metadata={"embedding_dimension": 1536},
+    ),
+    ProviderModelEntry(
+        provider_model_id="openai/text-embedding-3-large",
+        model_type=ModelType.embedding,
+        metadata={"embedding_dimension": 3072},
+    ),
+]

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

@@ -0,0 +1,22 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from llama_stack.providers.utils.inference.litellm_openai_mixin import LiteLLMOpenAIMixin
+
+from .config import OpenAIConfig
+from .models import MODEL_ENTRIES
+
+
+class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
+    def __init__(self, config: OpenAIConfig) -> None:
+        LiteLLMOpenAIMixin.__init__(self, MODEL_ENTRIES)
+        self.config = config
+
+    async def initialize(self) -> None:
+        pass
+
+    async def shutdown(self) -> None:
+        pass

llama_stack/providers/tests/test_cases/inference/chat_completion.json

@@ -40,7 +40,7 @@
   "tool_calling": {
     "data": {
       "messages": [
-        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "system", "content": "Pretend you are a weather assistant."},
         {"role": "user", "content": "What's the weather like in San Francisco?"}
       ],
       "tools": [
@@ -65,7 +65,7 @@
       "messages": [
         {
           "role": "system",
-          "content": "You are a helpful assistant."
+          "content": "Pretend you are a weather assistant."
         },
         {
           "role": "user",

llama_stack/providers/utils/inference/litellm_openai_mixin.py

@@ -0,0 +1,171 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from typing import AsyncGenerator, AsyncIterator, List, Optional, Union
+
+import litellm
+
+from llama_stack.apis.common.content_types import (
+    InterleavedContent,
+    InterleavedContentItem,
+)
+from llama_stack.apis.inference import (
+    ChatCompletionRequest,
+    ChatCompletionResponse,
+    ChatCompletionResponseStreamChunk,
+    EmbeddingsResponse,
+    EmbeddingTaskType,
+    Inference,
+    JsonSchemaResponseFormat,
+    LogProbConfig,
+    Message,
+    ResponseFormat,
+    SamplingParams,
+    TextTruncation,
+    ToolChoice,
+    ToolConfig,
+    ToolDefinition,
+    ToolPromptFormat,
+)
+from llama_stack.apis.models.models import Model
+from llama_stack.providers.utils.inference.model_registry import (
+    ModelRegistryHelper,
+)
+from llama_stack.providers.utils.inference.openai_compat import (
+    convert_message_to_openai_dict_new,
+    convert_openai_chat_completion_choice,
+    convert_openai_chat_completion_stream,
+    convert_tooldef_to_openai_tool,
+    get_sampling_options,
+)
+from llama_stack.providers.utils.inference.prompt_adapter import (
+    interleaved_content_as_str,
+)
+
+
+class LiteLLMOpenAIMixin(
+    ModelRegistryHelper,
+    Inference,
+):
+    def __init__(self, model_entries) -> None:
+        self.model_entries = model_entries
+        ModelRegistryHelper.__init__(self, model_entries)
+
+    async def register_model(self, model: Model) -> Model:
+        model_id = self.get_provider_model_id(model.provider_resource_id)
+        if model_id is None:
+            raise ValueError(f"Unsupported model: {model.provider_resource_id}")
+        return model
+
+    async def completion(
+        self,
+        model_id: str,
+        content: InterleavedContent,
+        sampling_params: Optional[SamplingParams] = SamplingParams(),
+        response_format: Optional[ResponseFormat] = None,
+        stream: Optional[bool] = False,
+        logprobs: Optional[LogProbConfig] = None,
+    ) -> AsyncGenerator:
+        raise NotImplementedError("LiteLLM does not support completion requests")
+
+    async def chat_completion(
+        self,
+        model_id: str,
+        messages: List[Message],
+        sampling_params: Optional[SamplingParams] = SamplingParams(),
+        tools: Optional[List[ToolDefinition]] = None,
+        tool_choice: Optional[ToolChoice] = ToolChoice.auto,
+        tool_prompt_format: Optional[ToolPromptFormat] = None,
+        response_format: Optional[ResponseFormat] = None,
+        stream: Optional[bool] = False,
+        logprobs: Optional[LogProbConfig] = None,
+        tool_config: Optional[ToolConfig] = None,
+    ) -> Union[ChatCompletionResponse, AsyncIterator[ChatCompletionResponseStreamChunk]]:
+        model = await self.model_store.get_model(model_id)
+        request = ChatCompletionRequest(
+            model=model.provider_resource_id,
+            messages=messages,
+            sampling_params=sampling_params,
+            tools=tools or [],
+            response_format=response_format,
+            stream=stream,
+            logprobs=logprobs,
+            tool_config=tool_config,
+        )
+
+        params = await self._get_params(request)
+
+        # unfortunately, we need to use synchronous litellm.completion here because litellm
+        # caches various httpx.client objects in a non-eventloop aware manner
+        response = litellm.completion(**params)
+        if stream:
+            return self._stream_chat_completion(response)
+        else:
+            return convert_openai_chat_completion_choice(response.choices[0])
+
+    async def _stream_chat_completion(
+        self, response: litellm.ModelResponse
+    ) -> AsyncIterator[ChatCompletionResponseStreamChunk]:
+        async def _stream_generator():
+            for chunk in response:
+                yield chunk
+
+        async for chunk in convert_openai_chat_completion_stream(
+            _stream_generator(), enable_incremental_tool_calls=True
+        ):
+            yield chunk
+
+    async def _get_params(self, request: ChatCompletionRequest) -> dict:
+        input_dict = {}
+
+        input_dict["messages"] = [await convert_message_to_openai_dict_new(m) for m in request.messages]
+        if fmt := request.response_format:
+            if not isinstance(fmt, JsonSchemaResponseFormat):
+                raise ValueError(
+                    f"Unsupported response format: {type(fmt)}. Only JsonSchemaResponseFormat is supported."
+                )
+
+            fmt = fmt.json_schema
+            name = fmt["title"]
+            del fmt["title"]
+            fmt["additionalProperties"] = False
+            input_dict["response_format"] = {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": name,
+                    "schema": fmt,
+                    "strict": True,
+                },
+            }
+        if request.tools:
+            input_dict["tools"] = [convert_tooldef_to_openai_tool(tool) for tool in request.tools]
+            if request.tool_config.tool_choice:
+                input_dict["tool_choice"] = request.tool_config.tool_choice.value
+
+        return {
+            "model": request.model,
+            **input_dict,
+            "stream": request.stream,
+            **get_sampling_options(request.sampling_params),
+        }
+
+    async def embeddings(
+        self,
+        model_id: str,
+        contents: List[str] | List[InterleavedContentItem],
+        text_truncation: Optional[TextTruncation] = TextTruncation.none,
+        output_dimension: Optional[int] = None,
+        task_type: Optional[EmbeddingTaskType] = None,
+    ) -> EmbeddingsResponse:
+        model = await self.model_store.get_model(model_id)
+
+        response = litellm.embedding(
+            model=model.provider_resource_id,
+            input=[interleaved_content_as_str(content) for content in contents],
+        )
+
+        embeddings = [data["embedding"] for data in response["data"]]
+        return EmbeddingsResponse(embeddings=embeddings)