DFlash Speculative Decoding Brings up to 15x LLM Inference Boost on NVIDIA Blackwell

Researchers at UC San Diego, in collaboration with NVIDIA, have released DFlash, an open-source block diffusion model for speculative decoding, designed to drastically improve large language model (LLM) inference performance on NVIDIA Blackwell GPUs. DFlash employs a block-diffusion drafter that generates an entire block of candidate tokens in a single forward pass, turning sequential drafting into parallel GPU work while preserving output quality. This approach delivers up to a 15x throughput improvement for gpt-oss-120b and significantly boosts interactivity for other models like Llama 3.1 8B, with minimal code changes required for integration into frameworks like vLLM and SGLang.