Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs

Flash-dLLM introduces IO-aware KV caching and parallel decoding to speed up inference and cut memory use in diffusion large language models (dLLMs), addressing the lack of effective Key-Value caching that has limited practical deployment of these non-autoregressive alternatives to autoregressive LLMs. The work targets the inefficient inference that currently constrains dLLM deployment.

Diffusion Large Language Models dLLMs have recently emerged as a promising alternative to autoregressive LLMs by enabling non-autoregressive text generation. However, their practical deployment remains limited by inefficient inference, largely due to the absence of effective Key-Value KV caching