---

name: nas-search description: "Routes AutoFormer, AutoFormerV2/S3, Cream, and CDARTS architecture-search, retrain, test, and sampled-ImageNet workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

NAS Search

Use this sub-skill when the user wants to search, retrain, evaluate, or debug one of the monorepo's architecture-search families: AutoFormer, AutoFormerV2/S3, Cream, or CDARTS.

What this route owns

• AutoFormer supernet training, evolution search, and evaluation.
• AutoFormerV2 / S3 evaluation from searched-space configs.
• Cream search, retrain, test, and subImageNet generation.
• CDARTS search, retrain, test, and the more advanced benchmark201 / detection / segmentation branches.

When to use it

Choose this route for prompts like:

• "search an AutoFormer architecture"
• "run Cream retraining"
• "evaluate S3 / AutoFormerV2"
• "fix CDARTS search config"
• "generate subImageNet"
• "compare NAS commands for these repos"

What to read next

• references/workflows.md for command shapes, config-file ownership, and which project uses which launcher.
• references/troubleshooting.md for legacy-torch, Apex, dataset-layout, and distributed-launch failures.
• scripts/build_nas_command.py to print safe command templates instead of launching a training job.
• ../../scripts/check_legacy_imports.py for the modern-torch compatibility shim used by legacy search code.
• ../../scripts/check_dataset_layout.py when the failure looks like a missing ImageNet, subImageNet, or COCO layout.

Important boundaries

• Do not route EfficientViT, MiniViT, TinyCLIP, TinyViT, or iRPE here.
• Treat Cream/tools/generate_subImageNet.py and AutoFormer/lib/subImageNet.py as reference evidence only; they mutate data and copy files, so the bundled skill uses read-only layout checks plus command templates instead.
• Treat CDARTS benchmark201 as an advanced optional path because it expects extra historical dependencies such as Apex and the NAS-Bench-201 API.

Working pattern

1. Identify the project family and the exact mode: train, search, retrain, test, or eval.
2. Read the workflow reference for the canonical command shape and the required config/data/checkpoint fields.
3. Run the bundled layout or compatibility checker if the error involves dataset paths, old torch imports, or legacy optional dependencies.
4. Use the command-builder script to produce the exact launcher string and then adapt it to the user's environment.

Common signals

• torch._six errors usually mean the legacy NAS code is being imported under modern torch.
• apex errors usually mean the request is trying to use the optional CDARTS benchmark201 path.
• FileNotFoundError on imagenet/train, imagenet/val, subImageNet, or coco usually means the data layout is wrong.
• Import-time parser errors usually mean the user tried to call the project script without the required mode/config arguments.