Conformal Sparsification for Bandwidth-Efficient Edge-Cloud Speculative Decoding

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Bhattacharjee, Payel, Tian, Fengwei, Zhong, Meiyu, Zhang, Guangyi, Simeone, Osvaldo, Tandon, Ravi
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909837277790208
author Bhattacharjee, Payel
Tian, Fengwei
Zhong, Meiyu
Zhang, Guangyi
Simeone, Osvaldo
Tandon, Ravi
author_facet Bhattacharjee, Payel
Tian, Fengwei
Zhong, Meiyu
Zhang, Guangyi
Simeone, Osvaldo
Tandon, Ravi
contents Edge-cloud speculative decoding (SD) accelerates inference by having a cloud-based large language model (LLM) that verifies draft tokens generated by a resource-constrained small language model (SLM) at the edge. A central bottleneck is the limited bandwidth of the edge-cloud link, which necessitates efficient compression of draft token distributions. We first derive an information-theoretic bound that decomposes the token rejection rate into contributions from SLM-LLM distribution mismatch and from quantization distortion. Guided by this analysis, we propose the Sparse Quantize-and-Sample SD (SQS-SD) framework, which exploits distributional sparsity through structured sparsification and lattice-based quantization. Within this framework, K-SQS applies fixed top-K truncation, while C-SQS adaptively adjusts the retained token set via online conformal prediction to ensure bounded deviation from the dense distribution. Empirical results confirm that both approaches improve end-to-end latency and rejection rates in complimentary operating regimes.
format Preprint
id arxiv_https___arxiv_org_abs_2510_09942
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Conformal Sparsification for Bandwidth-Efficient Edge-Cloud Speculative Decoding
Bhattacharjee, Payel
Tian, Fengwei
Zhong, Meiyu
Zhang, Guangyi
Simeone, Osvaldo
Tandon, Ravi
Machine Learning
Artificial Intelligence
Information Theory
Edge-cloud speculative decoding (SD) accelerates inference by having a cloud-based large language model (LLM) that verifies draft tokens generated by a resource-constrained small language model (SLM) at the edge. A central bottleneck is the limited bandwidth of the edge-cloud link, which necessitates efficient compression of draft token distributions. We first derive an information-theoretic bound that decomposes the token rejection rate into contributions from SLM-LLM distribution mismatch and from quantization distortion. Guided by this analysis, we propose the Sparse Quantize-and-Sample SD (SQS-SD) framework, which exploits distributional sparsity through structured sparsification and lattice-based quantization. Within this framework, K-SQS applies fixed top-K truncation, while C-SQS adaptively adjusts the retained token set via online conformal prediction to ensure bounded deviation from the dense distribution. Empirical results confirm that both approaches improve end-to-end latency and rejection rates in complimentary operating regimes.
title Conformal Sparsification for Bandwidth-Efficient Edge-Cloud Speculative Decoding
topic Machine Learning
Artificial Intelligence
Information Theory
url https://arxiv.org/abs/2510.09942