DFlash Accelerates LLM Inference up to 15x on NVIDIA Blackwell GPUs

Researchers at UC San Diego's z-lab released DFlash, a speculative decoding framework that significantly boosts large language model inference performance. DFlash introduces a block diffusion model that generates an entire block of candidate tokens in a single forward pass, replacing the slower autoregressive draft loop. NVIDIA independently confirmed that DFlash achieves up to 15 times the concurrent user load of standard autoregressive decoding on Blackwell hardware.