MAGPY: Compiling Eager Mode DNN Programs by Monitoring Execution States

Explore a groundbreaking approach to optimizing deep neural network (DNN) programs in this 19-minute conference talk from USENIX ATC '24. Delve into MAGPY, an innovative solution that addresses the challenges of compiling eager mode DNN programs written in dynamic programming languages like Python. Learn how MAGPY generates more complete operator graphs by monitoring program execution states, significantly improving performance. Discover the reference graph technique used to record execution states and identify critical state changes. Gain insights into experimental results showing up to 2.88× acceleration of complex deep learning programs and successful instantiation of 93.40% of real user programs into complete operator graphs. Understand the implications of this research for enhancing the efficiency of deep learning computations in real-world applications.