Opt4GPTQ: Co-Optimizing Memory and Computation for 4-bit GPTQ Quantized LLM Inference on Heterogeneous Platforms

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Yaozheng, Wang, Wei, Kong, Jie, Zhou, Jiehan, Zhang, Xianwei, Cui, Huanqing, Bao, Han, Liu, Yuhai
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912878457520128
author Zhang, Yaozheng
Wang, Wei
Kong, Jie
Zhou, Jiehan
Zhang, Xianwei
Cui, Huanqing
Bao, Han
Liu, Yuhai
author_facet Zhang, Yaozheng
Wang, Wei
Kong, Jie
Zhou, Jiehan
Zhang, Xianwei
Cui, Huanqing
Bao, Han
Liu, Yuhai
contents The increasing adoption of large language models (LLMs) on heterogeneous computing platforms poses significant challenges to achieving high inference efficiency. To address these efficiency bottlenecks across diverse platforms, this paper proposes Opt4GPTQ, a practical optimization method designed for 4-bit GPTQ quantized LLMs inference on heterogeneous AI accelerators. Built upon the vLLM serving system, Opt4GPTQ integrates three platform-level optimization strategies: Shared Memory Buffering Optimization (SMB-Opt), which caches frequently accessed data in shared memory and employs single-threaded writes; Vectorized Memory Loading Optimization (VML-Opt), which utilizes vectorized memory operations for efficient data loading; and Inline Assembly Optimization (ILA-Opt), which directly leverages hardwarenative vector half-precision addition and fused multiply-accumulate instructions. Experimental results show that Opt4GPTQ effectively improves performance across various models while maintaining original model accuracy, achieving throughput gains of up to 84.42%. This work highlights the critical role of platformlevel engineering in enabling efficient LLMs inference on emerging architectures and provides valuable methodologies for future heterogeneous platform adaptation.
format Preprint
id arxiv_https___arxiv_org_abs_2511_19438
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Opt4GPTQ: Co-Optimizing Memory and Computation for 4-bit GPTQ Quantized LLM Inference on Heterogeneous Platforms
Zhang, Yaozheng
Wang, Wei
Kong, Jie
Zhou, Jiehan
Zhang, Xianwei
Cui, Huanqing
Bao, Han
Liu, Yuhai
Distributed, Parallel, and Cluster Computing
Performance
The increasing adoption of large language models (LLMs) on heterogeneous computing platforms poses significant challenges to achieving high inference efficiency. To address these efficiency bottlenecks across diverse platforms, this paper proposes Opt4GPTQ, a practical optimization method designed for 4-bit GPTQ quantized LLMs inference on heterogeneous AI accelerators. Built upon the vLLM serving system, Opt4GPTQ integrates three platform-level optimization strategies: Shared Memory Buffering Optimization (SMB-Opt), which caches frequently accessed data in shared memory and employs single-threaded writes; Vectorized Memory Loading Optimization (VML-Opt), which utilizes vectorized memory operations for efficient data loading; and Inline Assembly Optimization (ILA-Opt), which directly leverages hardwarenative vector half-precision addition and fused multiply-accumulate instructions. Experimental results show that Opt4GPTQ effectively improves performance across various models while maintaining original model accuracy, achieving throughput gains of up to 84.42%. This work highlights the critical role of platformlevel engineering in enabling efficient LLMs inference on emerging architectures and provides valuable methodologies for future heterogeneous platform adaptation.
title Opt4GPTQ: Co-Optimizing Memory and Computation for 4-bit GPTQ Quantized LLM Inference on Heterogeneous Platforms
topic Distributed, Parallel, and Cluster Computing
Performance
url https://arxiv.org/abs/2511.19438