Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

Hugging Face integrates Nunchaku 4-bit diffusion inference into Diffusers, introducing Nunchaku Lite for native loading of quantized checkpoints without a separate inference engine, enabling faster inference and lower memory usage on compatible GPUs, including Blackwell-based NVIDIA GPUs for NVFP4 variants.

The integration uses the Diffusers library and the kernels package, making it simple to load pre-quantized models with from_pretrained().

Benchmarks show about 30% speedup and significant VRAM reduction compared to BF16 pipelines.

The diffuse-compressor toolkit allows users to quantize new architectures and publish them as standard Diffusers repositories.