DFlash Speculative Decoding Boosts Blackwell Throughput 15x

DFlash drafts entire blocks of tokens in parallel rather than one at a time, achieving up to 15x higher throughput on NVIDIA Blackwell GPUs. The method keeps output lossless by having a small draft model propose future tokens that a large target model verifies simultaneously. This addresses the serial bottleneck that leaves modern GPUs underused during autoregressive generation.