GraphMend: Code Transformations for Fixing Graph Breaks in PyTorch 2
The episode discusses GraphMend, a compiler technique from a University of Michigan team that automatically fixes 'graph breaks' in PyTorch 2, which force execution to slow Python eager mode. GraphMend uses source-level analysis and transformations on the Abstract Syntax Tree to ensure TorchDynamo captures continuous FX graphs. The research showed substantial performance gains, including up to twenty-six times cold-start speedup.