dLLM
dLLM: Simple Diffusion Language Modeling
...Unlike traditional autoregressive models that generate text sequentially token by token, diffusion language models generate text through an iterative denoising process that refines masked tokens over multiple steps. This approach allows models to reason over the entire sequence simultaneously and potentially produce more coherent outputs with bidirectional context. The project provides an integrated pipeline that standardizes how diffusion language models are trained, evaluated, and deployed, helping researchers reproduce experiments and compare results more easily. The framework includes scalable training infrastructure inspired by modern deep learning toolkits and supports integrations with widely used libraries for distributed training.