Duet Framework Enables Dual-Sparse Inference for Diffusion Language Models
November 9, 2026
Duet is a training-free inference framework that reduces the cost of diffusion language models through prefix KV sparsity and early query sparsity. It uses a compact historical KV subset reused across denoising steps and predicts update positions to avoid processing the entire block at every step.
HOW THIS AFFECTS YOU
●
builderYou can run diffusion-based language models more efficiently by reducing attention costs during generation.
●
researcherThis method provides a way to speed up inference without requiring additional training steps.