DFlash & DFlash-2
Diffusion-based speculative decoding for parallel multi-token drafting
Performance
In plain English
Instead of predicting one word after another like a slow typewriter, DFlash sketches an entire 5-word sentence block in parallel using diffusion and verifies it instantly.
Technically
Diffusion Speculative Flash Decoding (DFlash/DFlash-2). A non-autoregressive speculative drafting framework that generates multi-token candidate blocks via lightweight diffusion processes and verifies them in a single target transformer forward pass, eliminating sequential draft autoregression overhead.