Bringing Nunchaku 4-bit Diffusion Inference to Diffusers

За 30 секунд

Что произошло

Hugging Face integrated Nunchaku 4-bit quantization into Diffusers, enabling diffusion models to run with 4-bit weights and activations for faster inference and lower memory.

Почему это важно

Engineers deploying text-to-image models on consumer GPUs need to reduce VRAM from 20-30GB to fit models like ERNIE or FLUX on hardware with limited memory.

На что обратить внимание

NVFP4 kernels require NVIDIA Blackwell GPUs; earlier generations must use INT4 variants. Speedup is about 30 percent, less than the original Nunchaku engine's model-specific optimizations.