IBM describes native PyTorch integration for Spyre inference accelerator

IBM's torch-spyre team describes an integration that lets PyTorch recognize Spyre as a device and manage tensors stored on the accelerator. The design replaces a backend-specific runtime graph that left PyTorch treating tensors as CPU values, even when the runtime had moved them to Spyre. The integration keeps model graphs in PyTorch's compiler pipeline and prepares compiled programs once before submitting ordered operations. This reduces repeated launch work and supports device-resident tensors across operations. Data transfers can overlap computation, but independently submitted compute operations share one queue and execute sequentially.