Extending TorchInductor with FlyDSL: A New MLIR-Native Backend for High-Performance GEMMs

Сипаттама

Hello, I'm Liz Lee from AMD. >> Hello, I'm Ju from AMD. >> Modern AI workloads depends on a growing range of specialized GPU kernels and no single kernel generation back end is optimal across every operator ship and architecture. We explore this opportunity by integrating fly…