---

name: setup-and-configuration description: "Handles FastVideo installation, platform and backend selection, model registry lookup, presets, configuration schemas, and environment diagnosis." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Setup and configuration

Use for installation, Python/platform checks, CUDA/MPS/ROCm choice, model-ID support, preset selection, configuration parsing, and import or dependency failures.

Workflow

1. Identify OS, architecture, Python version, GPU/driver/toolkit, and whether the task needs inference, serving, preprocessing, training, or evaluation.
2. Choose an isolated Python 3.10–3.12 environment. NVIDIA installs select UV_TORCH_BACKEND=cu126 or cu130; Apple Silicon follows MPS guidance; ARM NVIDIA may require a source kernel build.
3. Install only the needed package extras. Do not install Linux CUDA-only extensions on MPS, and do not assume a CPU import proves CUDA support.
4. Run the bundled environment diagnostic.
5. Use configuration reference to build typed nested configs and model reference to check model families, workloads, and backend constraints.

The package exports VideoGenerator, PipelineConfig, SamplingParam, and __version__. The CLI exposes generate, serve, router-serve, bench, and eval.

Boundaries

Route generation flags and optimization choices to inference; route server config to serving; route data/training requirements to training-and-data. This route owns platform and config validation, not model-weight execution.

Checks

A successful import is only a package check. Also verify the selected torch backend, required optional modules, model-weight access, available VRAM, and whether the chosen model/optimization is registered. Read setup troubleshooting for recovery steps.