project-overview

Torch-spyre is an out-of-tree PyTorch backend for the IBM Spyre AI Accelerator. It provides an execution path for models using custom memory primitives called sticks and supports both eager and compiled execution modes.

The torch-spyre repository structure separates key files and directories for eager and compiled modes. You can navigate these distinct paths to understand the Inductor backend compilation pipeline and specific execution flows.

The Inductor backend compilation pipeline in torch-spyre utilizes key abstractions like SpyreTensorLayout and KernelSpec. These components define how operations are lowered and executed on the Spyre device architecture.

Yes, torch-spyre relies on important environment variables to control initialization flow and execution behavior. Configuring these variables is necessary for proper build system operation and testing on the Spyre device.

Torch-spyre implements specific testing strategies to validate both eager and compiled execution paths on the Spyre device. These strategies ensure the Inductor backend pipeline and custom abstractions function correctly.

Effectively using torch-spyre requires understanding the Spyre device architecture and its memory primitives, known as sticks. Key abstractions like SpyreTensorLayout depend on these architectural concepts for proper data handling.