← 研踪 PaperTrailarXiv:2611.25813 · v1
Efficient and Robust Drafting with Multi-Token Prediction
首次提交 2026-11-19 17:33
版本更新 2026-11-19 17:33
雷达观测 2026-11-20 19:58
摘要
Speculative decoding has emerged as a critical technique for accelerating Large Language Model (LLM) inference by decoupling generation into a drafting phase and a parallel verification phase. A crucial challenge is therefore designing a drafter that is both lightweight and accurate enough to approximate the predictions of the target model. We present a highly optimized Multi-Token Prediction (MTP) architecture that extends existing backbone LLMs to efficiently produce high-quality drafts. Built around a single shared drafting head, our approach features three core architectural optimizations: target activation injection to anchor semantic planning, zero-copy KV-cache cross-attention to bypass prefill compute and reduce memory overhead, and an efficient clustered vocabulary projection to reduce the cost of full-vocabulary decoding. Testing these improvements with the Gemma 4 model family, our G-MTP shows out-of-the-box speedups for both large server-side setups (up to 3.1) and smaller models targeting memory-constrained on-device environments (up to 2.3 on H100 and up to 2.2 on a smartphone GPU). In addition, we present parameter-efficient fine-tuning recipes for G-MTP that keep the drafter aligned with the backbone LLM under domain shift, requiring up to 1.83 fewer target forward passes than tuning only the backbone.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-19 17:33 · 最新观测版本
Efficient and Robust Drafting with Multi-Token Prediction
Speculative decoding has emerged as a critical technique for accelerating Large Language Model (LLM) inference by decoupling generation into a drafting phase and a parallel verification phase. A crucial challenge is therefore designing a drafter that is both lightweight and accurate enough to approximate the predictions of the target model. We present a highly optimized Multi-Token Prediction (MTP) architecture that extends existing backbone LLMs to efficiently produce high-quality drafts. Built around a single shared drafting head, our approach features three core architectural optimizations: target activation injection to anchor semantic planning, zero-copy KV-cache cross-attention to bypass prefill compute and reduce memory overhead, and an efficient clustered vocabulary projection to reduce the cost of full-vocabulary decoding. Testing these improvements with the Gemma 4 model family, our G-MTP shows out-of-the-box speedups for both large server-side setups (up to 3.1) and smaller models targeting memory-constrained on-device environments (up to 2.3 on H100 and up to 2.2 on a smartphone GPU). In addition, we present parameter-efficient fine-tuning recipes for G-MTP that keep the drafter aligned with the backbone LLM under domain shift, requiring up to 1.83 fewer target forward passes than tuning only the backbone.
查看此版本解读 →查看此版本原文 ↗