ES-dLLM: Efficient Inference for Diffusion Large Language Models by Early-Skipping

Researchers developed ES-dLLM, a training-free inference acceleration framework that speeds up diffusion large language models by selectively skipping tokens in early layers based on importance scoring. The method achieves 5.6x to 16.8x speedup over vanilla implementations while maintaining generation quality, offering a promising alternative to autoregressive models.

• →ES-dLLM delivers up to 16.8x speedup for diffusion large language models without requiring additional training
• →The framework achieves 226-308 tokens per second on NVIDIA H200 GPU while preserving generation quality
• →Token importance is computed using intermediate tensor variation and confidence scores from previous iterations
• →Diffusion LLMs show potential as alternatives to autoregressive models due to bidirectional context and parallel generation capabilities
• →The method outperforms state-of-the-art caching approaches by up to 1.85x in throughput improvements