Lightning Talk: Torch-Spyre: Compiling To a Multi-core Dataflow Accelerator... D. Grove & O. Tardieu

This talk covers the development of Torch-Spyre, an open source project designed to compile for the IBM Spyre Accelerator, which features 32 AI-optimized compute cores. David Grove and Olivier Tardieu from IBM describe how they evolved the Spyre software stack to leverage Inductor, resulting in reduced proprietary compiler code and improved compilation efficiency without sacrificing inference performance. The session highlights technical challenges faced in compiling for accelerators like Spyre and discusses adaptations made to Inductor, including support for device-specific tiled Tensor memory layouts and new optimization passes for effective scratchpad management. The speakers aim to engage the community to enhance support for these technologies within the PyTorch ecosystem.