sequential-diffusion-language-models

【数据与研究】Implement adaptive parallel decoding for language models using diffusion-based next-sequence prediction. Enable dynamic block-based token generation with confidence thresholds to achieve 2x+ speedups while maintaining competitive performance. Retrofit existing autoregressive models with minimal additional training data.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准