Shark-NLP/DiffuSeq

Implements classifier-free conditional diffusion for seq2seq tasks using PyTorch and HuggingFace transformers, with end-to-end training on dialogue, question generation, text simplification, and paraphrase datasets. DiffuSeq-v2 bridges discrete and continuous text spaces via learned soft absorbing states and discrete noise, achieving 4x faster convergence and 800x faster sampling via customized DPM-Solver++ integration. Supports distributed training across multiple GPUs with loss-aware schedule sampling and minimum Bayes risk decoding for quality/diversity trade-offs.

831 stars. No commits in the last 6 months.