Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models

Nemotron-Labs Diffusion introduces diffusion language models that generate multiple tokens in parallel and iteratively refine them, improving performance over traditional autoregressive generation.

• •Supports three modes: autoregressive for compatibility, diffusion for raw speed, self-speculation combining both
• •Achieves 2.6× tokens per forward pass in diffusion mode, up to 6.4× speedup in self-speculation mode
• •Trained on 1.3T tokens with joint AR and diffusion objective in 3B, 8B, 14B scales
• •Deployed through SGLang, allowing single-line config switching without application changes

This summary was automatically generated by AI based on the original article and may not be fully accurate.