arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2603.23911cs.CLcs.AIcs.LG
收藏

多令牌预测的自蒸馏

Self-Distillation for Multi-Token Prediction

Large Language Model Department, Tencent(腾讯大语言模型部门)

Guoliang Zhao, Ruobing Xie, An Wang, Shuaipeng Li, Huaibing Xie, Xingwu Sun

AI总结:

本文提出MTP-D自蒸馏方法,提升多令牌预测头接受率并保持主头性能,通过循环扩展策略进一步提升推理效率。

摘要:

随着大语言模型(LLMs)的扩展,推理效率成为关键瓶颈。多令牌预测(MTP)可通过并行预测多个未来令牌加速LLM推理。然而,现有MTP方法仍面临两个挑战:MTP头的接受率有限,以及难以联合训练多个MTP头。因此,我们提出MTP-D,一种简单有效的自蒸馏方法,具有最小的额外训练成本,可提升MTP头接受率(+7.5%)并最大限度保持主头性能。我们还引入了循环扩展策略,使MTP-D能够有效且经济地扩展MTP头,并进一步显著提升推理速度至1头MTP(+220.4%)。此外,我们通过在七个基准上的广泛实验系统探索并验证了蒸馏策略和MTP的潜在可扩展性。这些结果表明,MTP-D和循环扩展策略有效提升了MTP头性能和推理效率,促进了MTP在LLMs中的实际应用。

英文摘要:

As Large Language Models (LLMs) scale up, inference efficiency becomes a critical bottleneck. Multi-Token Prediction (MTP) could accelerate LLM inference by predicting multiple future tokens in parallel. However, existing MTP approaches still face two challenges: limited acceptance rates of MTP heads, and difficulties in jointly training multiple MTP heads. Therefore, we propose MTP-D, a simple yet effective self-distillation method with minimal additional training cost, which boosts MTP head acceptance rates (+7.5\%) while maximumly preserving main-head performance. We also introduce a looped extension strategy for MTP-D, enabling effective and economical MTP head extension and further significant inference speedup to 1-head MTP (+220.4\%). Moreover, we systematically explore and validate key insights on the distillation strategies and the potential scalability of MTP through extensive experiments on seven benchmarks. These results demonstrate that our MTP-D and looped extension strategy effectively enhance MTP-head performance and inference efficiency, facilitating the practical usage of MTP in LLMs.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑