DynaTE is a hardware-software co-design architecture for accelerating diffusion LLMs by dynamically adapting accelerator execution to evolving token states during decoding. This results in a 2.05-2.78x speedup and 2.99-3.93x higher energy efficiency compared to state-of-the-art dLLM accelerators.