Flash-dLLM is a training-free inference acceleration framework for diffusion large language models that identifies GPU memory I/O, rather than compute, as the dominant bottleneck when combining KV-cache reuse with parallel decoding. It introduces an I/O-aware fused KV-cache kernel to cut redundant memory movement and a draft-and-verify decoding strategy where the diffusion model serves as both drafter and verifier without an auxiliary model. On mathematical reasoning and code-generation benchmarks it achieves 5.1x and 11.0x speedups over the prior strongest baseline on GSM8K and HumanEval respectively.
