How would you use a diffusion model as the draft compon...

In a speculative decoding pipeline, a diffusion model can serve as a high-speed draft component. Because the diffusion model generates K tokens in a single forward pass, it avoids the sequential sampling bottleneck of an AR draft. The 70B AR target model then verifies the entire K-token sequence in parallel. To ensure efficiency, the diffusion draft must be fine-tuned on the target model’s output distribution to align the proposals, and the lookahead window K must be set as a fixed hyperparameter compatible with the diffusion model’s fixed-length constraint.