What problem does it solve?
This Skill helps PyTorch contributors understand, debug, and extend the Dynamo compiler without bypassing its symbolic execution, guard, or graph-generation architecture.
Core Features & Use Cases
- Dynamo Architecture Guidance: Explains frame interception, bytecode translation, VariableTrackers, FX graph assembly, backends, and the C++ runtime.
- Implementation Support: Provides patterns for adding opcode handlers, creating VariableTracker types, integrating pytrees, installing guards, handling side effects, and writing polyfills.
- Debugging Workflows: Diagnoses bytecode capture, object-to-variable conversion, reconstruction, graph breaks, guard failures, dynamic-shape issues, and compilation errors.
- Use Case: When a new Python type causes a torch.compile failure, use the Skill to identify the relevant VariableTracker, add builder support, implement reconstruction and method handling, create a regression test, and verify the fix.
Quick Start
Use the pytorch-dynamo skill to diagnose a torch.compile implementation failure and recommend the relevant code changes and tests.
