arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2405.16623cs.LGcs.ARcs.PF
收藏

用于张量编译器的具有配置交叉注意力的图神经网络

Graph neural networks with configuration cross-attention for tensor compilers

King Abdullah University of Science and Technology (KAUST)(阿卜杜拉国王科技大学)

Dmitrii Khizbullin, Eduardo Rocha de Andrade, Thanh Hau Nguyen, Matheus Pedroza Ferreira, David R. Pugh

AI总结:

本文提出神经图架构TGraph,通过配置交叉注意力筛选计算图中张量的快速转置与分块配置,将TpuGraphs布局集合上的平均Kendall's τ从29.8%提升至67.4%,实现AI张量编译器并显著降低推理能耗。

摘要:

随着神经网络近期的流行,对推理工作负载的高效服务需求日益增长。神经网络推理工作负载可以表示为一个计算图,其中节点是变换多维张量的算子。张量可以以组合上大量方式进行转置和/或分块,其中一些配置能够加速推理。我们提出TGraph,这是一种神经图架构,能够筛选目标计算图的快速配置,因此相对于传统的基于启发式的编译器,它代表了一种人工智能(AI)张量编译器。所提出的解决方案将TpuGraphs布局集合上的平均Kendall's τ从可靠基线的29.8%提高到TGraph的67.4%。我们估计,与我们的工作相关的潜在CO₂排放减少量相当于托管AI导向数据中心地区家庭总排放量的50%以上。

英文摘要:

With the recent popularity of neural networks comes the need for efficient serving of inference workloads. A neural network inference workload can be represented as a computational graph with nodes as operators transforming multidimensional tensors. The tensors can be transposed and/or tiled in a combinatorially large number of ways, some configurations leading to accelerated inference. We propose TGraph, a neural graph architecture that allows screening for fast configurations of the target computational graph, thus representing an artificial intelligence (AI) tensor compiler in contrast to the traditional heuristics-based compilers. The proposed solution improves mean Kendall's $τ$ across layout collections of TpuGraphs from 29.8% of the reliable baseline to 67.4% of TGraph. We estimate the potential CO$_2$ emission reduction associated with our work to be equivalent to over 50% of the total household emissions in the areas hosting AI-oriented data centers.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑