speeding up diffusion models with first block caching

how to speed up diffusion inference with minimal quality loss using first block caching