Torch-Spyre Documentation

Torch-Spyre is the PyTorch backend for the IBM Spyre AI Accelerator. It enables standard PyTorch models to run on the Spyre device with full torch.compile support via a custom Inductor backend.

New to Torch-Spyre?

Three on-ramps depending on what you need:

Just want to run a model? Start with Run PyTorch on Spyre device.
Need the mental model? Read Key concepts — a 5–10 minute primer on dataflow execution, sticks and tiled tensors, the LX scratchpad, the eager vs compiled paths, and graph breaks.
Want the design story? How Torch-Spyre works: an out-of-tree PyTorch backend walks through the four challenges we hit and the PyTorch extension mechanisms that addressed each one.

For a one-line definition of a specific term, jump to the Glossary.

For Users

For Developers

Torch-Spyre Documentation

Indices and tables