FlowBlock: Wavefront-Parallel Decoding for Self-Correcting Diffusion Language Models

2026-07-20

AI Summary

Block-level diffusion language models suffer from sequential inter-block dependencies that hinder parallel decoding and limit inference efficiency. This work proposes a training-free parallel decoding framework that transforms block-final states from hard dependencies into schedulable resources through a self-correction mechanism and wavefront scheduling strategy, enabling inter-block parallel generation while reusing KV caches. The approach integrates gated wavefront decoding, heterogeneous wavefront packing, token-to-token editing, and windowed block causal masking to support dense batching of asynchronous requests. Experiments demonstrate up to a 4.01× throughput improvement and 77.1% latency reduction over LLaDA-2.1/2.0, along with a 1.3-point average accuracy gain; compared to the trained baseline D2F, it achieves 16× higher batch throughput with superior precision.

Abstract

Block-wise diffusion large language models (dLLMs) decode sequentially at the block level, enabling effective KV-cache reuse across blocks but making inter-block decoding strictly serial. Prior work has attempted to unlock inter-block parallelism through post-training methods, but achieves only modest speedups and often degrades accuracy. We observe that self-correcting dLLMs offer a training-free alternative: token-to-token (T2T) editing can repair tokens drafted with a slightly stale upstream context, so a downstream block requires only an informative draft rather than a finalized predecessor. This turns block finality from a hard dependency into a scheduling resource. We propose \textbf{\flowblock{}}, a training-free parallel decoding framework built on two mechanisms. (i) \emph{Gated Wavefront Decoding} admits blocks into a bounded wavefront only when a readiness gate is satisfied, jointly refines active blocks via T2T editing, and commits blocks in order under a windowed block-causal mask that preserves exact frozen-prefix KV caches reuse. (ii) \emph{Heterogeneous Wavefront Packing} assigns each request an independent wavefront while packing asynchronous windows into dense, shape-stable batched forwards. Across different benchmarks, \flowblock{} improves tokens per second (TPS) over LLaDA-2.1 and LLaDA-2.0, two serial block-wise dLLMs, by up to 2.95$\times$ and 4.01$\times$, while reducing latency by up to 53.6\% and 77.1\%, respectively. It also improves average accuracy by 1.3 points. Compared with D2F, a training-based inter-block-parallel baseline, \flowblock{} achieves higher accuracy and up to 16$\times$ higher batched serving throughput.

Problem

• block-wise diffusion
• inter-block parallelism
• serial decoding
• KV-cache reuse
• decoding latency

Innovation

• FlowBlock
• wavefront-parallel decoding
• self-correcting diffusion language models
• KV-cache reuse
• training-free parallelism