---

name: utilities-and-configs description: "Operate Optimum utility APIs for dummy inputs, normalized configs, preprocessing processors, run configs, and save/load support." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Optimum utilities and configs

Use this sub-skill when a task needs Optimum support APIs rather than a backend-specific exporter, FX, or GPTQ workflow: dummy tensors for tracing/export probes, normalized access to heterogeneous Transformers configs, preprocessing task processors, benchmark/run config validation, and shared save/load base classes.

Operating map

Need	Use
Generate local dummy inputs for text, decoder, cache, bbox, vision, audio, labels, or diffusion-style inputs	references/dummy-inputs.md
Normalize nonstandard model configs and use NormalizedConfig.with_args	references/normalized-configs.md
Select task preprocessors, understand dataset caveats, validate run configs, or reason about base save/load classes	references/preprocessing-and-runs.md
Diagnose optional dependencies, task names, dtype/shape issues, and config local-vs-Hub errors	references/troubleshooting.md
Run a no-download local smoke check for normalized configs, text/vision dummy inputs, labels, and optional BaseConfig serialization	scripts/utils_smoke.py

Start here

1. Confirm the task is utility/config support. If the user asks for backend exporter registration, CLI routing, or accelerated pipeline backend setup, route to ../exporters-and-cli/SKILL.md. If they ask for FX graph transformations, route to ../fx-graph-workflows/SKILL.md. If they ask for GPTQ quantization workflows, route to ../gptq-quantization/SKILL.md.
2. For exporter or tracing dummy inputs, build or retrieve a normalized config first, then instantiate the narrowest DummyInputGenerator subclass that supports the model input name.
3. For unusual config attribute names, use NormalizedConfig.with_args(...) rather than monkey-patching the Transformers config.
4. Treat preprocessing processors and benchmark runs as optional-dependency surfaces: they can require torchvision, Pillow, datasets, model preprocessors, local datasets, or backend-specific subclasses.
5. Prefer local, bounded checks. Do not trigger model downloads, dataset downloads, training, quantization, Hub pushes, or destructive writes unless the user explicitly asks and provides a budget/credentials.

Fast verification

From this sub-skill directory or with an explicit path to the script:

python scripts/utils_smoke.py --help
python scripts/utils_smoke.py
python scripts/utils_smoke.py --framework np --skip-base-config
python scripts/utils_smoke.py --check-task-processors

The smoke script is intentionally no-download. It uses tiny in-memory config objects and a temporary directory for the optional BaseConfig round trip.