SD$^2$: Self-Distilled Sparse Drafters

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Lasby, Mike, Sinnadurai, Nish, Manohararajah, Valavan, Lie, Sean, Ioannou, Yani, Thangarasa, Vithursan
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915315089145856
author Lasby, Mike
Sinnadurai, Nish
Manohararajah, Valavan
Lie, Sean
Ioannou, Yani
Thangarasa, Vithursan
author_facet Lasby, Mike
Sinnadurai, Nish
Manohararajah, Valavan
Lie, Sean
Ioannou, Yani
Thangarasa, Vithursan
contents Speculative decoding is a powerful technique for reducing the latency of Large Language Models (LLMs), offering a fault-tolerant framework that enables the use of highly compressed draft models. In this work, we introduce Self-Distilled Sparse Drafters (SD$^2$), a novel methodology that leverages self-data distillation and fine-grained weight sparsity to produce highly efficient and well-aligned draft models. SD$^2$ systematically enhances draft token acceptance rates while significantly reducing Multiply-Accumulate operations (MACs), even in the Universal Assisted Generation (UAG) setting, where draft and target models originate from different model families. On a Llama-3.1-70B target model, SD$^2$ provides a 1.59$\times$ higher Mean Accepted Length (MAL) compared to layer-pruned draft models and reduces MACs by over 43.87% with a 8.36% reduction in MAL compared to a dense draft models. Our 1.5B and 3B unstructured sparse drafters outperform both dense and layer-pruned models in terms of end-to-end latency improvements; highlighting the potential of sparsity-aware fine-tuning and compression strategies to improve LLM inference efficiency while maintaining alignment with target models.
format Preprint
id arxiv_https___arxiv_org_abs_2504_08838
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SD$^2$: Self-Distilled Sparse Drafters
Lasby, Mike
Sinnadurai, Nish
Manohararajah, Valavan
Lie, Sean
Ioannou, Yani
Thangarasa, Vithursan
Computation and Language
Artificial Intelligence
I.2.7
Speculative decoding is a powerful technique for reducing the latency of Large Language Models (LLMs), offering a fault-tolerant framework that enables the use of highly compressed draft models. In this work, we introduce Self-Distilled Sparse Drafters (SD$^2$), a novel methodology that leverages self-data distillation and fine-grained weight sparsity to produce highly efficient and well-aligned draft models. SD$^2$ systematically enhances draft token acceptance rates while significantly reducing Multiply-Accumulate operations (MACs), even in the Universal Assisted Generation (UAG) setting, where draft and target models originate from different model families. On a Llama-3.1-70B target model, SD$^2$ provides a 1.59$\times$ higher Mean Accepted Length (MAL) compared to layer-pruned draft models and reduces MACs by over 43.87% with a 8.36% reduction in MAL compared to a dense draft models. Our 1.5B and 3B unstructured sparse drafters outperform both dense and layer-pruned models in terms of end-to-end latency improvements; highlighting the potential of sparsity-aware fine-tuning and compression strategies to improve LLM inference efficiency while maintaining alignment with target models.
title SD$^2$: Self-Distilled Sparse Drafters
topic Computation and Language
Artificial Intelligence
I.2.7
url https://arxiv.org/abs/2504.08838