---

name: "tuning-and-distributed" description: "Routes NeuralForecast Auto*, Ray, Optuna, and Spark distributed workflows." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

tuning-and-distributed

Use this sub-skill when the user wants Auto* hyperparameter search, backend resource configuration, or distributed Spark execution for NeuralForecast.

What this route covers

• BaseAuto and the Auto* wrappers.
• RayOptions and OptunaOptions.
• DistributedConfig and Spark-backed distributed paths.
• Resource, trial, and backend selection for tuning jobs.

What this route does not cover

• Panel dataframe preparation.
• Model-family choice beyond what is needed to size the search space.
• Loss mathematics.
• Core single-model fit/predict details unless they are needed to explain the tuning backend.

If the user only needs to run a model after the search is configured, route to ../core-forecasting/SKILL.md. If the user still needs to fix the panel or exogenous columns, route to ../data-and-exogenous/SKILL.md.

Read these bundled references

• ../../references/tuning-distributed.md for the backend matrix and workflow notes.
• ../../references/api-reference.md for the verified Auto*/backend signatures.
• references/troubleshooting.md for backend, resource, and search-space failures.

Run this bundled script

• ../../scripts/check_auto_config.py for a conservative Auto* configuration check.

Common triggers

• "I need AutoNHITS with Ray"
• "Can I use Optuna instead of Ray?"
• "How do I configure GPU trials?"
• "Can I run this with Spark?"
• "Why does cross-validation with use_fitted fail?"

Minimal route

1. Read ../../references/tuning-distributed.md.
2. Use ../../scripts/check_auto_config.py to confirm the backend option shape.
3. Read references/troubleshooting.md when a tuning or distributed job fails.

Decision cues

• If the issue is the model family itself, route to model-selection.
• If the issue is the panel dataframe or future exogenous layout, route to data-and-exogenous.
• If the user only wants a quick fit on a single model, route to core-forecasting.

Safe default

When in doubt, start with one tiny trial and the smallest plausible resource hint. Keep the configuration conservative until the search path is proven.