@inproceedings{liu-etal-2024-speculative-speculative,
title = "How Speculative Can Speculative Decoding Be?",
author = "Liu, Zhuorui and
Zhang, Chen and
Song, Dawei",
editor = "Calzolari, Nicoletta and
Kan, Min-Yen and
Hoste, Veronique and
Lenci, Alessandro and
Sakti, Sakriani and
Xue, Nianwen",
booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
month = may,
year = "2024",
address = "Torino, Italia",
publisher = "ELRA and ICCL",
url = "https://aclanthology.org/2024.lrec-main.725",
pages = "8265--8275",
abstract = "Large language models (LLMs) have drawn great attention from the field of natural language processing and beyond, due to their impressive capability of autoregressive modeling, yet bringing an obvious problem, i.e., the largely increased latency. An emerging idea to alleviate this problem is speculative decoding, which first uses a draft model to draft tokens autoregressively and then makes the target model verify these tokens in parallel. The draft model is typically smaller than the target model, and it essentially trades generation quality for speed. Thereby, speculative decoding can be viewed as a speculative game for the target model in term of verification failures. That is, the lengthy draft tokens proposed by the small draft models could fail in the verification stage. Naturally, a critical question arises: how speculative can speculative decoding be, or in other words, how small can an adequate draft model be and how large can an appropriate number of draft tokens be? This work aims to investigate these questions and demonstrate how the scale of the draft model and the number of draft tokens would have an impact on the overall latency of the speculative decoding. We theoretically show that neither of above two factors will be infinitely speculative. Namely, there is a certain turning point for each of them. We then empirically show that the scale of the draft model could be 10-20$\times$ smaller than the target model and the optimal number of draft tokens should lie in 3-5.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2024-speculative-speculative">
<titleInfo>
<title>How Speculative Can Speculative Decoding Be?</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhuorui</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chen</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dawei</namePart>
<namePart type="family">Song</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nicoletta</namePart>
<namePart type="family">Calzolari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Min-Yen</namePart>
<namePart type="family">Kan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Veronique</namePart>
<namePart type="family">Hoste</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="family">Lenci</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sakriani</namePart>
<namePart type="family">Sakti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nianwen</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA and ICCL</publisher>
<place>
<placeTerm type="text">Torino, Italia</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Large language models (LLMs) have drawn great attention from the field of natural language processing and beyond, due to their impressive capability of autoregressive modeling, yet bringing an obvious problem, i.e., the largely increased latency. An emerging idea to alleviate this problem is speculative decoding, which first uses a draft model to draft tokens autoregressively and then makes the target model verify these tokens in parallel. The draft model is typically smaller than the target model, and it essentially trades generation quality for speed. Thereby, speculative decoding can be viewed as a speculative game for the target model in term of verification failures. That is, the lengthy draft tokens proposed by the small draft models could fail in the verification stage. Naturally, a critical question arises: how speculative can speculative decoding be, or in other words, how small can an adequate draft model be and how large can an appropriate number of draft tokens be? This work aims to investigate these questions and demonstrate how the scale of the draft model and the number of draft tokens would have an impact on the overall latency of the speculative decoding. We theoretically show that neither of above two factors will be infinitely speculative. Namely, there is a certain turning point for each of them. We then empirically show that the scale of the draft model could be 10-20\times smaller than the target model and the optimal number of draft tokens should lie in 3-5.</abstract>
<identifier type="citekey">liu-etal-2024-speculative-speculative</identifier>
<location>
<url>https://aclanthology.org/2024.lrec-main.725</url>
</location>
<part>
<date>2024-05</date>
<extent unit="page">
<start>8265</start>
<end>8275</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T How Speculative Can Speculative Decoding Be?
%A Liu, Zhuorui
%A Zhang, Chen
%A Song, Dawei
%Y Calzolari, Nicoletta
%Y Kan, Min-Yen
%Y Hoste, Veronique
%Y Lenci, Alessandro
%Y Sakti, Sakriani
%Y Xue, Nianwen
%S Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)
%D 2024
%8 May
%I ELRA and ICCL
%C Torino, Italia
%F liu-etal-2024-speculative-speculative
%X Large language models (LLMs) have drawn great attention from the field of natural language processing and beyond, due to their impressive capability of autoregressive modeling, yet bringing an obvious problem, i.e., the largely increased latency. An emerging idea to alleviate this problem is speculative decoding, which first uses a draft model to draft tokens autoregressively and then makes the target model verify these tokens in parallel. The draft model is typically smaller than the target model, and it essentially trades generation quality for speed. Thereby, speculative decoding can be viewed as a speculative game for the target model in term of verification failures. That is, the lengthy draft tokens proposed by the small draft models could fail in the verification stage. Naturally, a critical question arises: how speculative can speculative decoding be, or in other words, how small can an adequate draft model be and how large can an appropriate number of draft tokens be? This work aims to investigate these questions and demonstrate how the scale of the draft model and the number of draft tokens would have an impact on the overall latency of the speculative decoding. We theoretically show that neither of above two factors will be infinitely speculative. Namely, there is a certain turning point for each of them. We then empirically show that the scale of the draft model could be 10-20\times smaller than the target model and the optimal number of draft tokens should lie in 3-5.
%U https://aclanthology.org/2024.lrec-main.725
%P 8265-8275
Markdown (Informal)
[How Speculative Can Speculative Decoding Be?](https://aclanthology.org/2024.lrec-main.725) (Liu et al., LREC-COLING 2024)
ACL
- Zhuorui Liu, Chen Zhang, and Dawei Song. 2024. How Speculative Can Speculative Decoding Be?. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 8265–8275, Torino, Italia. ELRA and ICCL.