A field guide for ML engineers, LLM DevOps, and system architects deploying open-weights 30B-scale models with block-diffusion speculative decoding.
Introduction & Background
1.1 What DFlash actually is
DFlash ("Block Diffusion for Flash Speculative Decoding") is a speculative decoding framework out of UC San Diego's z-lab that replaces the autoregressive drafter used in methods like EAGLE-2/EAGLE-3 with a lightweight block-diffusion drafter. Instead of generating draft tokens one at a time (itself a sequential process, just cheaper than the target model), DFlash's drafter is conditioned on hidden-state features pulled from the target model and denoises an entire block of candidate tokens typically 8 to 16 in a single forward pass. The target model then verifies the whole block in parallel, exactly as in classic speculative decoding.
Two things make this architecturally distinct from EAGLE-style drafters:
