Skip to content

Add a reproducible NeMo Gym example for comparing routing strategies #559

Description

@cwing-nvidia

Motivation

Switchyard’s existing benchmark example shows how to evaluate routing with Harbor and coding-agent benchmarks. Switchyard also integrates with NeMo Gym, which supports a broader collection of benchmarks and agent harnesses, but the Switchyard repository does not currently include a runnable example of that workflow.

Adding one would help users:

  • Evaluate routing strategies on benchmarks beyond Harbor’s coding-agent workloads.
  • Compare a routed strategy with a fixed-model baseline using the same tasks.
  • Reproduce the workflow directly from the Switchyard repository.
  • Discover and validate the NeMo Gym integration without assembling commands from external documentation.

Proposed solution

Add benchmark/nemo_gym/ with a small, executable, version-pinned example that:

  • Defines a fixed-model baseline and one routed strategy.
  • Runs both against the same benchmark and task subset.
  • Writes each condition to a separate output directory.
  • Pairs rollouts before comparing rewards.
  • Compares model tokens, classifier tokens, latency, and routing statistics.
  • Explains the generated provenance and statistics artifacts.
  • Links to the full NeMo Gym integration documentation.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions