Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2511.21759 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912731975647232 |
|---|---|
| author | Wei, Linye Chen, Wenjue Tang, Pingzhi Guo, Xiaotian Ye, Le Wang, Runsheng Li, Meng |
| author_facet | Wei, Linye Chen, Wenjue Tang, Pingzhi Guo, Xiaotian Ye, Le Wang, Runsheng Li, Meng |
| contents | Diffusion-based large language models (dLLMs) have recently gained significant attention for their exceptional performance and inherent potential for parallel decoding. Existing frameworks further enhance its inference efficiency by enabling KV caching. However, its bidirectional attention mechanism necessitates periodic cache refreshes that interleave prefill and decoding phases, both contributing substantial inference cost and constraining achievable speedup. Inspired by the heterogeneous arithmetic intensity of the prefill and decoding phases, we propose ODB-dLLM, a framework that orchestrates dual-boundaries to accelerate dLLM inference. In the prefill phase, we find that the predefined fixed response length introduces heavy yet redundant computational overhead, which affects efficiency. To alleviate this, ODB-dLLM incorporates an adaptive length prediction mechanism that progressively reduces prefill overhead and unnecessary computation. In the decoding phase, we analyze the computational characteristics of dLLMs and propose a dLLM-specific jump-share speculative decoding method to enhance efficiency by reducing the number of decoding iterations. Experimental results demonstrate that ODB-dLLM achieves 46-162x and 2.63-6.30x speedups over the baseline dLLM and Fast-dLLM, respectively, while simultaneously mitigating the accuracy degradation in existing acceleration frameworks. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2511_21759 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Orchestrating Dual-Boundaries: An Arithmetic Intensity Inspired Acceleration Framework for Diffusion Language Models Wei, Linye Chen, Wenjue Tang, Pingzhi Guo, Xiaotian Ye, Le Wang, Runsheng Li, Meng Computation and Language Machine Learning Diffusion-based large language models (dLLMs) have recently gained significant attention for their exceptional performance and inherent potential for parallel decoding. Existing frameworks further enhance its inference efficiency by enabling KV caching. However, its bidirectional attention mechanism necessitates periodic cache refreshes that interleave prefill and decoding phases, both contributing substantial inference cost and constraining achievable speedup. Inspired by the heterogeneous arithmetic intensity of the prefill and decoding phases, we propose ODB-dLLM, a framework that orchestrates dual-boundaries to accelerate dLLM inference. In the prefill phase, we find that the predefined fixed response length introduces heavy yet redundant computational overhead, which affects efficiency. To alleviate this, ODB-dLLM incorporates an adaptive length prediction mechanism that progressively reduces prefill overhead and unnecessary computation. In the decoding phase, we analyze the computational characteristics of dLLMs and propose a dLLM-specific jump-share speculative decoding method to enhance efficiency by reducing the number of decoding iterations. Experimental results demonstrate that ODB-dLLM achieves 46-162x and 2.63-6.30x speedups over the baseline dLLM and Fast-dLLM, respectively, while simultaneously mitigating the accuracy degradation in existing acceleration frameworks. |
| title | Orchestrating Dual-Boundaries: An Arithmetic Intensity Inspired Acceleration Framework for Diffusion Language Models |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2511.21759 |