NVIDIA Safety Provider Calling Wrong Guardrails Endpoint

[Hadar301]

System Info

• LlamaStack Version: 0.4.0.dev0
• Distribution: nvidia
• Provider: remote::nvidia safety provider
• Guardrails Service: NeMo Guardrails 0.10.x

ISSUE-nvidia-safety-provider-bug.md

Information

• [x] The official example scripts
• [ ] My own modified scripts

🐛 Describe the bug

The NVIDIA safety provider in LlamaStack is calling the wrong endpoint when communicating with NeMo Guardrails service, causing safety/shield functionality to fail with 500 Internal Server Error.

Steps to Reproduce

1. Configure LlamaStack with nvidia safety provider:

providers:
  safety:
    - provider_id: nvidia
      provider_type: remote::nvidia
      config:
        guardrails_service_url: http://nemoguardrails-sample:8000
        config_id: demo-self-check-input-output
        model: meta/llama-3.2-1b-instruct

2. Register a shield:

curl -X POST http://localhost:8321/v1/shields \
  -H "Content-Type: application/json" \
  -d '{
    "shield_id": "demo-self-check-input-output",
    "provider_id": "nvidia",
    "provider_shield_id": "demo-self-check-input-output",
    "params": {"model": "meta/llama-3.2-1b-instruct"}
  }'

3. Create guardrails config in NeMo Guardrails service:

curl -X POST http://guardrails-service:8000/v1/guardrail/configs \
  -H "Content-Type: application/json" \
  -d '{
    "name": "demo-self-check-input-output",
    "namespace": "default",
    "data": {
      "prompts": [...],
      "rails": {...}
    }
  }'

4. Try to run shield via LlamaStack API:

curl -X POST http://localhost:8321/v1/safety/run-shield \
  -H "Content-Type: application/json" \
  -d '{
    "shield_id": "demo-self-check-input-output",
    "messages": [{"role": "user", "content": "You are stupid"}]
  }'

Error logs

ERROR 2025-11-19 15:18:45,931 llama_stack.core.server.server:285 core::server: Error executing endpoint
route='/v1/safety/run-shield' method='post'
╭───────────────────────────────────── Traceback (most recent call last) ─────────────────────────────────────╮ │ /workspace/src/llama_stack/core/server/server.py:275 in route_handler │ │ │ │ 272 │ │ │ │ │ return StreamingResponse(gen, media_type="text/event-stream") │ │ 273 │ │ │ │ else: │ │ 274 │ │ │ │ │ value = func(**kwargs) │ │ ❱ 275 │ │ │ │ │ result = await maybe_await(value) │ │ 276 │ │ │ │ │ if isinstance(result, PaginatedResponse) and result.url is None: │ │ 277 │ │ │ │ │ │ result.url = route │ │ 278 │ │ │ │ /workspace/src/llama_stack/core/server/server.py:197 in maybe_await │ │ │ │ 194 │ │ 195 async def maybe_await(value): │ │ 196 │ if inspect.iscoroutine(value): │ │ ❱ 197 │ │ return await value │ │ 198 │ return value │ │ 199 │ │ 200 │ │ │ │ /workspace/src/llama_stack/core/telemetry/trace_protocol.py:103 in async_wrapper │ │ │ │ 100 │ │ │ │ │ 101 │ │ │ with tracing.span(f"{class_name}.{method_name}", span_attributes) as span: │ │ 102 │ │ │ │ try: │ │ ❱ 103 │ │ │ │ │ result = await method(self, *args, **kwargs) │ │ 104 │ │ │ │ │ span.set_attribute("output", serialize_value(result)) │ │ 105 │ │ │ │ │ return result │ │ 106 │ │ │ │ except Exception as e: │ │ │ │ /workspace/src/llama_stack/core/routers/safety.py:60 in run_shield │ │ │ │ 57 │ ) -> RunShieldResponse: │ │ 58 │ │ logger.debug(f"SafetyRouter.run_shield: {shield_id}") │ │ 59 │ │ provider = await self.routing_table.get_provider_impl(shield_id) │ │ ❱ 60 │ │ return await provider.run_shield( │ │ 61 │ │ │ shield_id=shield_id, │ │ 62 │ │ │ messages=messages, │ │ 63 │ │ │ params=params, │ │ │ │ /workspace/src/llama_stack/core/telemetry/trace_protocol.py:103 in async_wrapper │ │ │ │ 100 │ │ │ │ │ 101 │ │ │ with tracing.span(f"{class_name}.{method_name}", span_attributes) as span: │ │ 102 │ │ │ │ try: │ │ ❱ 103 │ │ │ │ │ result = await method(self, *args, **kwargs) │ │ 104 │ │ │ │ │ span.set_attribute("output", serialize_value(result)) │ │ 105 │ │ │ │ │ return result │ │ 106 │ │ │ │ except Exception as e: │ │ │ │ /workspace/src/llama_stack/providers/remote/safety/nvidia/nvidia.py:67 in run_shield │ │ │ │ 64 │ │ │ raise ValueError(f"Shield {shield_id} not found") │ │ 65 │ │ │ │ 66 │ │ self.shield = NeMoGuardrails(self.config, shield.shield_id) │ │ ❱ 67 │ │ return await self.shield.run(messages) │ │ 68 │ │ │ 69 │ async def run_moderation(self, input: str | list[str], model: str | None = None) -> │ │ ModerationObject: │ │ 70 │ │ raise NotImplementedError("NVIDIA safety provider currently does not implement │ │ run_moderation") │ │ │ │ /workspace/src/llama_stack/providers/remote/safety/nvidia/nvidia.py:147 in run │ │ │ │ 144 │ │ │ │ "config_id": self.config_id, │ │ 145 │ │ │ }, │ │ 146 │ │ } │ │ ❱ 147 │ │ response = await self._guardrails_post(path="/v1/guardrail/checks", │ │ data=request_data) │ │ 148 │ │ │ │ 149 │ │ if response["status"] == "blocked": │ │ 150 │ │ │ user_message = "Sorry I cannot do this." │ │ │ │ /workspace/src/llama_stack/providers/remote/safety/nvidia/nvidia.py:117 in _guardrails_post │ │ │ │ 114 │ │ │ "Accept": "application/json", │ │ 115 │ │ } │ │ 116 │ │ response = requests.post(url=f"{self.guardrails_service_url}{path}", │ │ headers=headers, json=data) │ │ ❱ 117 │ │ response.raise_for_status() │ │ 118 │ │ return response.json() │ │ 119 │ │ │ 120 │ async def run(self, messages: list[OpenAIMessageParam]) -> RunShieldResponse: │ │ │ │ /usr/local/lib/python3.12/site-packages/requests/models.py:1026 in raise_for_status │ │ │ │ 1023 │ │ │ ) │ │ 1024 │ │ │ │ 1025 │ │ if http_error_msg: │ │ ❱ 1026 │ │ │ raise HTTPError(http_error_msg, response=self) │ │ 1027 │ │ │ 1028 │ def close(self): │ │ 1029 │ │ """Releases the connection back to the pool. Once this method has been │ ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ HTTPError: 500 Server Error: Internal Server Error for url:
http://nemoguardrails-sample.hacohen-nemo.svc.cluster.local:8000/v1/guardrail/checks

Expected behavior

The safety provider should successfully communicate with the NeMo Guardrails service and return a safety response indicating whether the content should be blocked.