mirror of https://github.com/meta-llama/llama-stack.git synced 2025-08-16 06:27:58 +00:00

Ashwin Bharambe cc87995e2b

chore: rename templates to distributions (#3035 )

As the title says. Distributions is in, Templates is out.

`llama stack build --template` --> `llama stack build --distro`. For
backward compatibility, the previous option is kept but results in a
warning.

Updated `server.py` to remove the "config_or_template" backward
compatibility since it has been a couple releases since that change.

2025-08-04 11:34:17 -07:00

1.7 KiB

Raw Permalink Blame History

NVIDIA Safety Provider for LlamaStack

This provider enables safety checks and guardrails for LLM interactions using NVIDIA's NeMo Guardrails service.

Features

Run safety checks for messages

Getting Started

Prerequisites

LlamaStack with NVIDIA configuration
Access to NVIDIA NeMo Guardrails service
NIM for model to use for safety check is deployed

Setup

Build the NVIDIA environment:

llama stack build --distro nvidia --image-type venv

Basic Usage using the LlamaStack Python Client

Initialize the client

import os

os.environ["NVIDIA_API_KEY"] = "your-api-key"
os.environ["NVIDIA_GUARDRAILS_URL"] = "http://guardrails.test"

from llama_stack.core.library_client import LlamaStackAsLibraryClient

client = LlamaStackAsLibraryClient("nvidia")
client.initialize()

Create a safety shield

from llama_stack.apis.safety import Shield
from llama_stack.apis.inference import Message

# Create a safety shield
shield = Shield(
    shield_id="your-shield-id",
    provider_resource_id="safety-model-id",  # The model to use for safety checks
    description="Safety checks for content moderation",
)

# Register the shield
await client.safety.register_shield(shield)

Run safety checks

# Messages to check
messages = [Message(role="user", content="Your message to check")]

# Run safety check
response = await client.safety.run_shield(
    shield_id="your-shield-id",
    messages=messages,
)

# Check for violations
if response.violation:
    print(f"Safety violation detected: {response.violation.user_message}")
    print(f"Violation level: {response.violation.violation_level}")
    print(f"Metadata: {response.violation.metadata}")
else:
    print("No safety violations detected")

1.7 KiB Raw Permalink Blame History