Open-source scheduler turns PyTorch work mappings into IBM Spyre pipelines

A new open-source Dataflow Scheduler converts PyTorch compiler work mappings into execution pipelines for IBM's Spyre accelerator, according to a PyTorch technical post. The scheduler takes KTIR, a representation of which cores perform computations and where their data resides, and organizes loads, computation and stores through the hardware's memory hierarchy. The scheduler and its intermediate representations are available under Apache-2.0. Its output feeds device-specific compiler stages that produce code and a job plan for repeated runtime launches. The scheduler preserves the frontend's work division; device-specific code generation starts after its output boundary.