DFlash Speculative Decoding: Parallel Token Blocks Boost Throughput Up to 15x on NVIDIA Blackwell
Decision Brief
What changedUC San Diego's DFlash replaces autoregressive draft generation with a lightweight block diffusion model, generating entire token blocks in a single forward pass for accelerated speculative decoding.
Why it mattersDFlash introduces a novel speculative decoding approach that significantly boosts inference throughput and supports mainstream inference engines, offering valuable insights for AI builders in model deployment and optimization.
Who should careAll AI builders, Inference / infra teams
Affected stackQwenNVIDIA
Builder actionMonitor
Source confidenceMedium · Reliable media or first-hand reporting
DFlash is a speculative decoding technique from UC San Diego. Its core innovation replaces traditional autoregressive draft generation with a lightweight block diffusion model, enabling parallel generation of entire token blocks in a single forward pass and conditioning target hidden features via KV injection. The paper reports up to 6.08x lossless speedup on Qwen3-8B, while NVIDIA achieves 15x throughput improvement on Blackwell architecture under fixed interactivity. DFlash has released 20 checkpoints and supports mainstream inference engines including SGLang, vLLM, and TensorRT-LLM.
Summary basis: official / RSS sourceUnless it says 'full article read', this summary is based only on publicly available content — it never pretends to have read restricted originals.
Sources
- MarkTechPost
Fast research-paper and ML tooling summaries, useful for infra and agent updates.
- MarkTechPost