Security for NeMo Platform

View as Markdown

This page provides security guidelines and best practices for deploying and managing NeMo Platform in production environments.

Security Considerations

  • The NeMo Platform does not impose rate limits. You must implement a rate-limiting strategy to restrict access to your application.
  • The NeMo Platform does not have an internal notion of a user. To restrict authorization to specific endpoints or users, implement an external mechanism such as an Envoy proxy.
  • The NeMo Data Store microservice does not provide object-class-specific access controls. All items reside within a single access control boundary.
  • The NeMo Platform, by design, can access all content in the NeMo Data Store microservice, including LoRA adapters, training data, evaluation data, and evaluation results. This access is required for model evaluation. Carefully weigh the risk of data exposure from serving customized models directly to production against the overhead of a separate deployment.
  • The NeMo Platform is not intended to be internet-facing. Deploy them as the logic (middle) tier in a three-tier architecture.
  • You are responsible for securing access to any application using the microservices. This includes:
  • Implementing an authentication layer between users and your application
  • Applying required authorization controls
  • Securing communication between services in your application

Refer to the NVIDIA Product Security page for information about subscribing to bulletins and updates, managing vulnerabilities, and reporting vulnerabilities.

Default Network Ports

The following table lists the default network ports for each microservice or default database. You can override these port numbers during deployment.

Network PortMicroservice
443/TCPNeMo Admission Service API
3000/TCPNeMo Data Store API
7331/TCPNeMo Evaluator API
7331/TCPNeMo Guardrails API
8000/TCPNIM API
8000/TCPNeMo Retriever Text Embedding API
8000/TCPNeMo Retriever Text Reranking API
8000/TCPNeMo Customizer API
8000/TCPNeMo Entity Store API
8443/TCPNeMo Operator metrics
8080/TCPVolcano Scheduler metrics
9009/TCPNeMo Customizer callback

By default, the NeMo Platform Helm Chart configures databases with the following network ports. Alternatively, you can configure each microservice to use an external database during installation.

Network PortDatabase
5432/TCPNeMo Customizer Database
5432/TCPNeMo Data Store Database
5432/TCPNeMo Evaluator Database
5432/TCPNeMo Entity Store Database
9091/TCPMilvus metrics
19530/TCPMilvus API

By default, the NeMo Platform Helm Chart installs an open telemetry collector, which uses the following network ports:

  • 4317/TCP
  • 4318/TCP
  • 6831/UDP
  • 9411/TCP
  • 14250/TCP
  • 14268/TCP

Refer to the OpenTelemetry Documentation for more information.