Reproduce “Dynamic-dLLM: Dynamic Cache-Budget and Adaptive Parallel Decoding for Training-Free Acceleration of Diffusion LLM” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tianyi Wu, Xiaoxi Sun, Yanhua Jiao, Yulin Li, Yixin Chen, YunHao Cao, YiQi Hu, Zhuotao Tian
DYNAMIC-DLLM, a training-free framework from Harbin Institute of Technology, Shenzhen and Huawei, accelerates Diffusion Large Language Models (dLLMs) by adaptively managing cache updates and parallel decoding. The framework achieved an average inference speedup of over 3x and up to 4.48x across LLaDA and Dream models on diverse benchmarks, while preserving generation quality.
Key insights
- Token properties, such as internal feature dynamics and prediction confidence, vary significantly across different model layers and decoding steps, highlighting the inefficiency of static acceleration methods.
- Layer input changes are a reliable and computationally inexpensive proxy for identifying dynamic intermediate features, enabling adaptive allocation of cache update resources.
- Combining prediction confidence concentration with temporal instability provides an effective mechanism for dynamically adjusting unmasking thresholds, allowing for earlier commitment to stable predictions without compromising accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.