Model WarsJune 24, 2026via MarkTechPost
DFlash Speculative Decoding Drafts Whole Token Blocks in Parallel for Up to 15x Higher Throughput on NVIDIA Blackwell
Why it matters
DFlash introduces a fundamental shift in speculative decoding architecture, moving from token-by-token drafting to parallel block generation. This unlocks measurable throughput gains on production inference infrastructure (Blackwell), directly impacting deployment economics for inference-heavy applications.
Key signals
- 6.08x lossless speedup on Qwen3-8B (paper results)
- Up to 15x throughput on NVIDIA Blackwell (at fixed interactivity)
- Block diffusion model replaces autoregressive drafting
- KV injection conditions on target hidden features
- 20 checkpoints shipped
- Integrations: SGLang, vLLM, TensorRT-LLM
- UC San Diego research
The hook
15x throughput. That's what DFlash achieves on NVIDIA Blackwell by replacing autoregressive drafting with parallel block diffusion—and it's shipping to SGLang, vLLM, and TensorRT-LLM today.
UC San Diego's DFlash replaces autoregressive drafting with a lightweight block diffusion model for speculative decoding. It drafts whole token blocks in a single forward pass and conditions on target hidden features through KV injection. The paper reports up to 6.08x lossless speedup on Qwen3-8B, while NVIDIA reports up to 15x throughput on Blackwell at fixed interactivity. DFlash ships 20 checkpoints and supports SGLang, vLLM, and TensorRT-LLM.