Flash-dLLM addresses the inefficient inference of Diffusion Large Language Models (dLLMs) by introducing an IO-aware approach to Key-Value (KV) caching and parallel decoding. Unlike existing methods that treat caching and decoding in isolation, this work identifies and mitigates the I/O bottlenecks arising from cache reuse during non-autoregressive text generation. By jointly optimizing both components, Flash-dLLM achieves faster and more memory-efficient diffusion LLM inference. The paper demonstrates that integrating IO-awareness into both KV caching and parallel decoding significantly improves practical deployability of dLLMs.