Robots Atlas>ROBOTS ATLAS
Inference

DFlash

2026ResearchPublished: 19 November 2026Updated: 19 November 2026Published
Key innovation
Replaces sequential, autoregressive drafting in speculative decoding with a lightweight block-diffusion model that generates an entire block of draft tokens in a single forward pass.
Category
Inference
Abstraction level
Technika
Operation level
InferenceServing
Use cases
Accelerating large language model inferenceReducing generation latency in real-time applications (chatbots, assistants)Increasing throughput and GPU utilization when serving LLMsLossless decoding acceleration without changing target model output quality

How it works

1) A lightweight block-diffusion draft model generates a whole block of candidate draft tokens in a single forward pass instead of producing them sequentially. 2) The drafter is conditioned on context features extracted from the target model, improving draft accuracy. 3) The target model verifies the block of draft tokens in parallel. 4) Accepted tokens are kept and the process stays lossless — the target model output distribution is preserved.

Problem solved

Autoregressive LLMs decode sequentially, token by token, causing high inference latency and poor GPU utilization. Existing speculative decoding mitigates this, but the draft model itself still runs autoregressively and sequentially, which limits practical speedups.

Key mechanisms

Block-diffusion drafting — generating a block of draft tokens in a single forward pass
Conditioning the drafter on context features from the target model
Parallel verification of draft tokens by the target model
Lossless token acceptance preserving the target model distribution

Strengths & limitations

Strengths
✓Over 6x lossless acceleration across a range of models and tasks (per the originating paper)
✓Up to 2.5x higher speedup than the SOTA EAGLE-3 (per the originating paper)
✓Parallel drafting removes the sequential bottleneck of autoregressive drafters
✓Higher acceptance rate thanks to conditioning on target-model features
✓Lossless — does not change target model output quality
Limitations
✗Requires training a separate block-diffusion draft model aligned to the target model
✗Actual speedup depends on the acceptance rate, which varies across models and tasks
✗Diffusion models historically underperform autoregressive models in quality — DFlash mitigates this via conditioning on target-model features

Components

Block-diffusion draft modelDrafter

A lightweight block-diffusion model that generates an entire block of draft tokens in a single forward pass rather than sequentially.

Official

Context-feature conditioningConditioning

A mechanism conditioning the drafter on context features extracted from the target model, increasing draft quality and the acceptance rate.

Target model verifierVerifier

The target model (large LLM) that verifies the block of draft tokens in parallel and guarantees lossless generation.

Evolution

Original paper · 2026 · ICML 2026 · Jian Chen
DFlash: Block Diffusion for Flash Speculative Decoding
Jian Chen, Yesheng Liang, Zhijian Liu
2026
DFlash introduced
Inflection point

Introduction of DFlash — speculative decoding with a lightweight block-diffusion drafter; accepted at ICML 2026, with public code.

Computational complexity

Computational characteristics
→Drafting: a single forward pass of the block-diffusion model (parallel generation of a token block)
→Verification: a parallel forward pass of the target model over the draft token block
→Goal: improve GPU utilization and reduce the latency of memory-bound autoregressive decoding
Benchmark notes

Per the originating paper, DFlash achieves over 6x lossless acceleration across a range of models and tasks and up to 2.5x higher speedup than the SOTA EAGLE-3 method. Per-model/per-task figures are in the paper (arXiv:2602.06036, ICML 2026).

Parallelism

Parallelism level
Fully parallel

Drafting generates a whole block of tokens in parallel in a single forward pass, and target-model verification runs in parallel over the entire block.

Scope
InferenceAcross tokens

Hardware requirements

Primary

The method targets improved GPU utilization and reduced LLM inference latency; parallel drafting and verification exploit GPU accelerator parallelism.