Graph neural networks with configuration cross-attention for tensor compilers

Dmitrii Khizbullin^1*Eduardo Rocha De Andrade²Thanh Hau Nguyen²Matheus Pedroza Ferreira²David Robert Pugh¹

• ¹King Abdullah University of Science and Technology, Thuwal, Saudi Arabia
• ²Sprout AI, London, United Kingdom

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 traditional heuristic-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.