Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models

In three linesPrefilling-dLLM optimizes diffusion language model inference by partitioning context into chunks, caching their KV representations, and selecting relevant chunks with intra-chunk token sparsity. Achieves 9.1–28.0x speedup on 8K–32K contexts without full prefix re-encoding.

Summary generated by Claude — human-verified