Overview

nunchaku_lite is organized around a small public API and model-specific adapters. The usual workflow is:

1. Start from a standard Diffusers pipeline and model id.
2. Load a Nunchaku SVDQ checkpoint with load_nunchaku_pipeline(...).
3. Use the resulting Diffusers pipeline normally for prompting, scheduling, and runtime LoRA loading.

The documentation in this section covers:

• Benchmarks: measured latency, peak CUDA memory, transformer storage, charts, and generated samples.
• Supported models: runnable model-specific loading guides.
• API Reference: public loading, patching, and adapter registry APIs.
• Roadmap: supported model coverage and remaining feature work.
• Documentation Deployment: docs update and publishing flow.