Getting Started

This section introduces Torch-Spyre, walks you through installing it, and shows how to run your first PyTorch model on the Spyre device.

• How Torch-Spyre works: an out-of-tree PyTorch backend
  • A device with a different execution model from a GPU
  • Challenge 1: making PyTorch recognize a new device
  • Challenge 2: teaching PyTorch a memory layout it had never seen
  • Challenge 3: extending TorchInductor for dataflow compilation
  • Challenge 4: covering ops in a model forward pass
  • What we learned
  • What is next
  • Getting started
  • Appendix: extension point reference for out-of-tree PyTorch backends
  • Acknowledgments
• Key concepts
  • 1. Execution model
  • 2. Hardware
  • 3. Memory hierarchy
  • 4. Sticks and tiled tensors
  • 5. Eager vs compiled path
  • 6. Graph breaks
  • 7. Compilation pipeline
  • 8. Dtype defaults and casting
  • 9. Running models today: FMS vs stock HuggingFace
  • 10. Hardware constraints checklist
  • 11. Streams
  • 12. Distributed execution
  • Where to go next
• Glossary
• Installation
  • Status
  • Requirements
  • Verify the Installation
  • Running the Test Suite
  • Next Steps
• Run PyTorch on Spyre device
  • Creating a Tensor
  • Running Tensor Operations
• More Examples