nas

NNI NAS

Use this sub-skill when the task is about NNI Neural Architecture Search: defining a model space, choosing mutation primitives, selecting an evaluator and strategy, launching a NasExperiment, exporting a fixed architecture, or diagnosing NAS import/runtime dependency failures.

Route Here

• The user asks to build or adapt an NNI NAS ModelSpace, including DARTS-style cells, LayerChoice, InputChoice, Repeat, Cell, nni.choice, or mutable torch.nn modules.
• The user asks whether to use one-shot NAS or multi-trial NAS, or asks about Random, GridSearch, RegularizedEvolution, TPE, PolicyBasedRL, DARTS, ENAS, GumbelDARTS, RandomOneShot, or Proxyless.
• The user asks for evaluator selection, FunctionalEvaluator, PyTorch Lightning evaluators, Classification, Regression, Lightning, LightningModule, or serialization with nni.trace.
• The user asks how to export, freeze, retrain, or instantiate the best architecture after a NAS search using export_top_models, model_context, or freeze.
• The user hits torch, pytorch_lightning, nni.nas.nn.pytorch, nni.nas.evaluator, or nni.nas.strategy import errors while doing NAS work.

Reroute

• Generic HPO experiment lifecycle, nnictl, search spaces for tuners, experiment YAML, training services, and trial-code wiring: use ../hpo-experiments/.
• Pruning, quantization, config_list, pruner, quantizer, speedup, or compression evaluators: use ../model-compression/.
• Standalone tracing, serialization utilities, concrete trace inspection, or feature utility usage without a NAS goal: use ../feature-engineering-and-utilities/.

Start With These References

• references/workflows.md for model-space, evaluator, strategy, experiment, and fixed-architecture workflows.
• references/api-reference.md for key public NAS concepts and object names, including optional dependency boundaries.
• references/troubleshooting.md for backend imports, evaluator mismatch, benchmark data, GPU/training cost, and model-space mutation errors.
• scripts/check_nas_optional_deps.py to safely inspect local optional dependency/import readiness without running NAS examples.

Default Approach

1. Confirm the task is NAS, not generic HPO or compression.
2. Separate the NAS design into model space, evaluator, strategy, execution configuration, and export/retrain plan.
3. Prefer minimal, conceptual scaffolds unless the user explicitly asks to run a search; NAS execution can be expensive and often requires torch, datasets, Lightning, GPUs, and long training time.
4. Warn when a chosen strategy requires optional dependencies that may not be installed; do not imply torch or pytorch_lightning are present by default.
5. For fixed-architecture handoff, export an architecture dict and show how to instantiate or freeze the final model from that dict.