longllada-diffusion-context

Training-free method extending diffusion LLMs to 6x context length using NTK-based RoPE scaling, exploiting bidirectional attention stability.

Independent GitHub snapshot · Automated classification · Original documentation governs