mit-han-lab/lpd

Introduces **Flexible Parallelized Autoregressive Modeling** with learnable position query tokens enabling arbitrary generation ordering and mutual visibility among concurrent tokens, paired with **Locality-aware Generation Ordering** that minimizes dependencies within groups while maximizing contextual support. Reduces generation steps from 256 to 20 (256×256) and 1024 to 48 (512×512) on ImageNet class-conditional generation, achieving 3.4x+ latency improvements over prior parallelized autoregressive approaches. Built on discrete tokenization (LlamaGen VQ-GAN) with pre-trained models from 337M to 1.4B parameters available on Hugging Face.