Diffusion Draft Model
A model that formulates draft generation as a diffusion denoising process and can produce multiple draft tokens in parallel within a single forward pass. It is used to accelerate speculative decoding and overcomes the one-by-one generation limitation of autoregressive draft models.
Timeline (last 90 days)
In diffusion draft model training, when the query budget is fixed, the acceptance length remains stable as the context size grows; simply increasing the query budget yields diminishing quality gains.
Diffusion draft models can denoise multiple draft tokens in parallel in a single forward pass, making them an important approach for accelerating speculative decoding.
All Facts (2)
In diffusion draft model training, when the query budget is fixed, the acceptance length remains stable as the context size grows; simply increasing the query budget yields diminishing quality gains.
50%UnverifiedDiffusion draft models can denoise multiple draft tokens in parallel in a single forward pass, making them an important approach for accelerating speculative decoding.
50%