generator: NeMoGuardrails server support

[jmartin-tech]

Add support for new OpenAICompatible service exposed via NVIDIA-NeMo/Guardrails

Verification

List the steps needed to make sure this thing works

• Supporting configuration file

plugins:
  target_type: guardrails.NeMoGuardrailsServer
  target_name: meta/llama-3.3-70b-instruct
  generators:
    guardrails:
      NeMoGuardrailsServer:
        uri: http://localhost:8005/v1/
        config_ids:
          - abc

• start a NeMoGuardrails service example used is abc bot with a revised backing model and default config_id:

NVIDIA_API_KEY="my super secret" MAIN_MODEL_ENGINE='nim' nemoguardrails server --config abc/ --port 8000 --default-config-id abc

• garak -t guardrails.NeMoGuardrailsServer -n meta/llama-3.3-70b-instruct -p lmrc

• Verify the target guardrails configuration responds

• start a NeMoGuardrails service example used is abc bot with a revised backing model without a default config_id:

NVIDIA_API_KEY="my super secret" MAIN_MODEL_ENGINE='nim' nemoguardrails server --config abc/ --port 8005

• garak --config config.guardrails.yaml -p lmrc
• Verify the target guardrails abc configuration is activated and responds

[leondz]

couple of requests for clarification

[leondz]

Is this still a sensible default?

[jmartin-tech]

I would be in favor of shifting if there is consensus. That would elevate this to a breaking change for this addition as configurations that did not specify -t guardrails.NeMoGuardrails would change behavior.

[leondz]

is it worth setting ENV_VAR/key_env_var to None, or deleting ENV_VAR, to reduce potential confusion from this attr being inherited?

[jmartin-tech]

Interesting point, currently NeMo Guardrails server does not offer any built in authentication, leading to the suppression added here is matched to the public docs from the sdk. I agree there is some confusion created by keeping the ENV_VAR on the class. I will test other patterns for this suppression.

[jmartin-tech]

ENV_VAR is inherited so I added support for it being None as a suppression of the requirement in Configurable.

A place holder value still needs to be set here to ensure the openai client library does not raise when no key is supplied or found in their standard variable locations. Should we adjust this to apply the place holder after the call to super() to allow a user to provide api_key or key_env_var explicitly in the config file for this generator? This might future proof this better if NeMo Guardrails were to ever add support for injecting an authentication layer on the the hosting server.

[leondz]

this is mysterious to me. i accept it, and welcome the mystery.

-- is something being worked around here? maybe a brief explanation in the docstring / a comment

[jmartin-tech]

Official docs this was based on here. Will work out how to document this better on the class.

[jmartin-tech]

Please validate update in d531028 meets the ask here.