FlowBlock
A training-free, wavefront-parallel decoding framework for block-wise diffusion language models letting a downstream block start from an informative draft instead of waiting on its predecessor, reporting higher throughput at similar accuracy on unmodified LLaDA-2 weights.
FlowBlock is a training-free parallel decoding framework for block-wise diffusion language models such as the LLaDA-2 series, which normally decode one block at a time, making inter-block decoding strictly serial. It exploits the token-to-token self-correction ability of the LLaDA-2.1 series so a downstream block can start from an informative draft rather than waiting for a finalized predecessor, treating block finality as a scheduling resource rather than a hard dependency.
It combines two mechanisms: Gated Wavefront Decoding admits blocks into a bounded wavefront only once a readiness gate is satisfied, jointly refines active blocks through token-to-token editing, and commits blocks in order under a windowed block-causal mask that preserves exact frozen-prefix KV-cache reuse; Heterogeneous Wavefront Packing gives each request its own wavefront while packing all active windows into dense, shape-stable batched forward passes so asynchronous progress stays efficient on a GPU under batched serving.
It reports up to about three times higher throughput than LLaDA-2.1 and four times higher than LLaDA-2.0 under batched serving, with lower latency and an average accuracy improvement across eight math and code benchmarks, running the unmodified LLaDA-2.1 series weights with no additional training. It requires Linux, an NVIDIA GPU with about 80 GB of memory for the 16B-A1B mixture-of-experts checkpoints, and a CUDA 12.x driver.