Transferable Graph Optimizers for ML Compilers

TL;DR

This paper introduces a transferable deep reinforcement learning approach for graph optimization in ML compilers, achieving significant speedups and improvements over existing methods and human experts across various complex graph tasks.

Contribution

It presents a novel end-to-end, scalable graph neural network-based reinforcement learning method that generalizes to unseen graphs and optimizes entire computational graphs efficiently.

Findings

33%-60% speedup on three optimization tasks

21% average improvement over human experts

18% improvement over prior state-of-the-art methods

Abstract

Most compilers for machine learning (ML) frameworks need to solve many correlated optimization problems to generate efficient machine code. Current ML compilers rely on heuristics based algorithms to solve these optimization problems one at a time. However, this approach is not only hard to maintain but often leads to sub-optimal solutions especially for newer model architectures. Existing learning based approaches in the literature are sample inefficient, tackle a single optimization problem, and do not generalize to unseen graphs making them infeasible to be deployed in practice. To address these limitations, we propose an end-to-end, transferable deep reinforcement learning method for computational graph optimization (GO), based on a scalable sequential attention mechanism over an inductive graph neural network. GO generates decisions on the entire graph rather than on each individual node autoregressively, drastically speeding up the search compared to prior methods. Moreover, we propose recurrent attention layers to jointly optimize dependent graph optimization tasks and demonstrate 33%-60% speedup on three graph optimization tasks compared to TensorFlow default optimization. On a diverse set of representative graphs consisting of up to 80,000 nodes, including Inception-v3, Transformer-XL, and WaveNet, GO achieves on average 21% improvement over human experts and 18% improvement over the prior state of the art with 15x faster convergence, on a device placement task evaluated in real systems.