Fast Multistage Compilation of Machine Learning Computation Graphs

Abstract

Machine learning applications are increasingly requiring fast and more computational power. Many applications like language models have become so large that they are run on distributed systems in parallel. However, getting into the details of optimally scheduling or even just running machine learning models on distributed systems can be a distraction for researchers ideating models. Hence there has been development of abstractions to facilitate running machine learning models in parallel on distributed systems. We present a compiler for the StreamIt language- a language made for abstract signal processing and multicore programming. We use that abstraction as a way to distribute the computation of machine learning models programmed in PyTorch.