Draft Model Knows When to Stop: A Self-Verification Length Policy for Speculative Decoding
Ziyin Zhang, Jiahao Xu, Tian Liang, Xingyu Chen, Zhiwei He, Rui Wang, Zhaopeng Tu
2024-11-28
Summary
This paper presents a new method called SVIP that improves the efficiency of Speculative Decoding (SD) in large language models by using a dynamic draft length based on the difficulty of generating tokens.
What's the problem?
Speculative Decoding is a technique that speeds up how language models generate text, but traditional methods use a fixed length for draft tokens, which can be inefficient. This means they might not adapt well to different tasks, leading to slower performance and wasted resources.
What's the solution?
The authors introduce SVIP, a method that adjusts the length of draft sequences based on how difficult it is to generate each token. By analyzing the complexity of the task, SVIP can determine the optimal number of tokens to draft, which helps speed up the overall process without needing to retrain the model or change its structure. This approach allows for faster and more efficient text generation.
Why it matters?
This research is important because it enhances the performance of large language models, making them quicker and more efficient in generating text. This improvement can benefit various applications, such as chatbots, content creation, and any system that relies on fast and accurate language processing.
Abstract
Speculative Decoding (SD) has become an important technique in accelerating the inference speed of large language models. Conventional SD methods employ a fixed draft length, which ignores the token generation difficulty across tasks. Consequently, in this paper, we address such an issue and introduce SVIP - a difficulty-aware dynamic draft length policy for speculative decoding systems. Based on a theoretical lower bound of draft token acceptance rate and its inference-time approximation, SVIP adaptively determines the lengths of draft sequences based on the entropy of each draft token distribution. Experimental results on mainstream SD benchmarks and frameworks demonstrate the superior performance of SVIP, achieving up to 20\% walltime speedup on SpecBench over baseline SD methods and 60\% speedup on MT-Bench for long-form generation of up to 8K tokens. Moreover, SVIP is totally training-free and compatible with any existing SD methods that generate draft tokens autoregressively. Experimental results also show that SVIP yields consistent walltime improvement on top of GliDe & CaPE and EAGLE-2.