Unlocking the Full Potential of TorchDynamo: Accelerating, Comparing, and Debugging ML Systems
October 16, 2026 Yi Pan — UC Berkeley
Abstract
TorchDynamo is best known as the frontend of torch.compile, but its ability to capture and decouple computation graphs remains underused. This talk presents our group’s recent work that builds on it to tackle three AI infrastructure bottlenecks. For acceleration, DynaFlow separates model definition from physical execution to add intra-device parallelism without invasive rewrites, and Piper enables flexible inter- and intra-device parallelism for complex schedules such as DualPipe. For differential comparison, Magneton uses graph tracing to match equivalent operators across systems and identify the implementation with lower latency or energy use. For numerical debugging, we extend the comparison to debug numerical correctness by capturing and comparing intermediate tensors at matched subgraph boundaries.
Speaker Bio
Yi Pan is a first-year PhD student in the Sky Computing Lab at UC Berkeley, advised by Prof. Ion Stoica. His research interests lie in the intersection of computer systems and AI. Currently, he is interested in designing and studying efficient systems and model architectures for large-scale LLM training and inference. His work has appeared at top-tier system conferences including MLSys, NSDI, and EuroSys. Prior to joining Berkeley, Yi was a research intern at the UW SyFI Lab.