feat: added nvidia as safety provider (#1248)

# What does this PR do?
Adds nvidia as a safety provider by interfacing with the nemo guardrails
microservice.
This enables checking user’s input or the LLM’s output against input and
output guardrails by using the `/v1/guardrails/checks` endpoint of the[
guardrails
API.](https://developer.nvidia.com/docs/nemo-microservices/guardrails/source/guides/checks-guide.html)

## Test Plan
Deploy nemo guardrails service following the documentation:
https://developer.nvidia.com/docs/nemo-microservices/guardrails/source/getting-started/deploy-docker.html

### Standalone:
```bash
(venv) local-cdgamarose@a1u1g-rome-0153:~/llama-stack$ pytest -v -s llama_stack/providers/tests/safety/test_safety.py --providers inference=nvidia,safety=nvidia --safety-shield meta/llama-3.1-8b-instruct

=================================================================================== test session starts ===================================================================================
platform linux -- Python 3.10.12, pytest-8.3.4, pluggy-1.5.0 -- /localhome/local-cdgamarose/llama-stack/venv/bin/python3
cachedir: .pytest_cache
metadata: {'Python': '3.10.12', 'Platform': 'Linux-5.15.0-122-generic-x86_64-with-glibc2.35', 'Packages': {'pytest': '8.3.4', 'pluggy': '1.5.0'}, 'Plugins': {'metadata': '3.1.1', 'asyncio': '0.25.3', 'anyio': '4.8.0', 'html': '4.1.1'}}
rootdir: /localhome/local-cdgamarose/llama-stack
configfile: pyproject.toml
plugins: metadata-3.1.1, asyncio-0.25.3, anyio-4.8.0, html-4.1.1
asyncio: mode=strict, asyncio_default_fixture_loop_scope=None
collected 2 items

llama_stack/providers/tests/safety/test_safety.py::TestSafety::test_shield_list[--inference=nvidia:safety=nvidia] Initializing NVIDIASafetyAdapter(http://0.0.0.0:7331)...
PASSED
llama_stack/providers/tests/safety/test_safety.py::TestSafety::test_run_shield[--inference=nvidia:safety=nvidia] PASSED

============================================================================== 2 passed, 2 warnings in 4.78s ==============================================================================

```
### Distribution:
```
llama stack run llama_stack/templates/nvidia/run-with-safety.yaml
curl -v -X 'POST' "http://localhost:8321/v1/safety/run-shield" -H 'accept: application/json' -H 'Content-Type: application/json' -d '{"shield_id": "meta/llama-3.1-8b-instruct", "messages":[{"role": "user", "content": "you are stupid"}]}'
{"violation":{"violation_level":"error","user_message":"Sorry I cannot do this.","metadata":{"self check input":{"status":"blocked"}}}}
```

[//]: # (## Documentation)

---------

Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>

llama_stack/providers/registry/safety.py

@@ -55,4 +55,13 @@ def available_providers() -> List[ProviderSpec]:
                 config_class="llama_stack.providers.remote.safety.bedrock.BedrockSafetyConfig",
             ),
         ),
+        remote_provider_spec(
+            api=Api.safety,
+            adapter=AdapterSpec(
+                adapter_type="nvidia",
+                pip_packages=["requests"],
+                module="llama_stack.providers.remote.safety.nvidia",
+                config_class="llama_stack.providers.remote.safety.nvidia.NVIDIASafetyConfig",
+            ),
+        ),
     ]

llama_stack/providers/remote/safety/nvidia/__init__.py

@@ -0,0 +1,18 @@
+# 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
+
+from .config import NVIDIASafetyConfig
+
+
+async def get_adapter_impl(config: NVIDIASafetyConfig, _deps) -> Any:
+    from .nvidia import NVIDIASafetyAdapter
+
+    impl = NVIDIASafetyAdapter(config)
+    await impl.initialize()
+    return impl

llama_stack/providers/remote/safety/nvidia/config.py

@@ -0,0 +1,37 @@
+# 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.
+import os
+from typing import Any, Dict, Optional
+
+from pydantic import BaseModel, Field
+
+from llama_stack.schema_utils import json_schema_type
+
+
+@json_schema_type
+class NVIDIASafetyConfig(BaseModel):
+    """
+    Configuration for the NVIDIA Guardrail microservice endpoint.
+
+    Attributes:
+        guardrails_service_url (str): A base url for accessing the NVIDIA guardrail endpoint, e.g. http://0.0.0.0:7331
+        config_id (str): The ID of the guardrails configuration to use from the configuration store
+         (https://developer.nvidia.com/docs/nemo-microservices/guardrails/source/guides/configuration-store-guide.html)
+
+    """
+
+    guardrails_service_url: str = Field(
+        default_factory=lambda: os.getenv("GUARDRAILS_SERVICE_URL", "http://0.0.0.0:7331"),
+        description="The url for accessing the guardrails service",
+    )
+    config_id: Optional[str] = Field(default="self-check", description="Config ID to use from the config store")
+
+    @classmethod
+    def sample_run_config(cls, **kwargs) -> Dict[str, Any]:
+        return {
+            "guardrails_service_url": "${env.GUARDRAILS_SERVICE_URL:http://localhost:7331}",
+            "config_id": "self-check",
+        }

llama_stack/providers/remote/safety/nvidia/nvidia.py

@@ -0,0 +1,154 @@
+# 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.
+
+import logging
+from typing import Any, List, Optional
+
+import requests
+
+from llama_stack.apis.inference import Message
+from llama_stack.apis.safety import RunShieldResponse, Safety, SafetyViolation, ViolationLevel
+from llama_stack.apis.shields import Shield
+from llama_stack.distribution.library_client import convert_pydantic_to_json_value
+from llama_stack.providers.datatypes import ShieldsProtocolPrivate
+
+from .config import NVIDIASafetyConfig
+
+logger = logging.getLogger(__name__)
+
+
+class NVIDIASafetyAdapter(Safety, ShieldsProtocolPrivate):
+    def __init__(self, config: NVIDIASafetyConfig) -> None:
+        """
+        Initialize the NVIDIASafetyAdapter with a given safety configuration.
+
+        Args:
+            config (NVIDIASafetyConfig): The configuration containing the guardrails service URL and config ID.
+        """
+        print(f"Initializing NVIDIASafetyAdapter({config.guardrails_service_url})...")
+        self.config = config
+
+    async def initialize(self) -> None:
+        pass
+
+    async def shutdown(self) -> None:
+        pass
+
+    async def register_shield(self, shield: Shield) -> None:
+        if not shield.provider_resource_id:
+            raise ValueError("Shield model not provided.")
+
+    async def run_shield(
+        self, shield_id: str, messages: List[Message], params: Optional[dict[str, Any]] = None
+    ) -> RunShieldResponse:
+        """
+        Run a safety shield check against the provided messages.
+
+        Args:
+            shield_id (str): The unique identifier for the shield to be used.
+            messages (List[Message]): A list of Message objects representing the conversation history.
+            params (Optional[dict[str, Any]]): Additional parameters for the shield check.
+
+        Returns:
+            RunShieldResponse: The response containing safety violation details if any.
+
+        Raises:
+            ValueError: If the shield with the provided shield_id is not found.
+        """
+        shield = await self.shield_store.get_shield(shield_id)
+        if not shield:
+            raise ValueError(f"Shield {shield_id} not found")
+
+        self.shield = NeMoGuardrails(self.config, shield.shield_id)
+        return await self.shield.run(messages)
+
+
+class NeMoGuardrails:
+    """
+    A class that encapsulates NVIDIA's guardrails safety logic.
+
+    Sends messages to the guardrails service and interprets the response to determine
+    if a safety violation has occurred.
+    """
+
+    def __init__(
+        self,
+        config: NVIDIASafetyConfig,
+        model: str,
+        threshold: float = 0.9,
+        temperature: float = 1.0,
+    ):
+        """
+        Initialize a NeMoGuardrails instance with the provided parameters.
+
+        Args:
+            config (NVIDIASafetyConfig): The safety configuration containing the config ID and guardrails URL.
+            model (str): The identifier or name of the model to be used for safety checks.
+            threshold (float, optional): The threshold for flagging violations. Defaults to 0.9.
+            temperature (float, optional): The temperature setting for the underlying model. Must be greater than 0. Defaults to 1.0.
+
+        Raises:
+            ValueError: If temperature is less than or equal to 0.
+            AssertionError: If config_id is not provided in the configuration.
+        """
+        self.config_id = config.config_id
+        self.model = model
+        assert self.config_id is not None, "Must provide config id"
+        if temperature <= 0:
+            raise ValueError("Temperature must be greater than 0")
+
+        self.temperature = temperature
+        self.threshold = threshold
+        self.guardrails_service_url = config.guardrails_service_url
+
+    async def run(self, messages: List[Message]) -> RunShieldResponse:
+        """
+        Queries the /v1/guardrails/checks endpoint of the NeMo guardrails deployed API.
+
+        Args:
+            messages (List[Message]): A list of Message objects to be checked for safety violations.
+
+        Returns:
+            RunShieldResponse: If the response indicates a violation ("blocked" status), returns a
+            RunShieldResponse with a SafetyViolation; otherwise, returns a RunShieldResponse with violation set to None.
+
+        Raises:
+            requests.HTTPError: If the POST request fails.
+        """
+        headers = {
+            "Accept": "application/json",
+        }
+        request_data = {
+            "model": self.model,
+            "messages": convert_pydantic_to_json_value(messages),
+            "temperature": self.temperature,
+            "top_p": 1,
+            "frequency_penalty": 0,
+            "presence_penalty": 0,
+            "max_tokens": 160,
+            "stream": False,
+            "guardrails": {
+                "config_id": self.config_id,
+            },
+        }
+        response = requests.post(
+            url=f"{self.guardrails_service_url}/v1/guardrail/checks", headers=headers, json=request_data
+        )
+        response.raise_for_status()
+        if "Content-Type" in response.headers and response.headers["Content-Type"].startswith("application/json"):
+            response_json = response.json()
+        if response_json["status"] == "blocked":
+            user_message = "Sorry I cannot do this."
+            metadata = response_json["rails_status"]
+
+            return RunShieldResponse(
+                violation=SafetyViolation(
+                    user_message=user_message,
+                    violation_level=ViolationLevel.ERROR,
+                    metadata=metadata,
+                )
+            )
+        return RunShieldResponse(violation=None)