Aoti Debug

What it does

The aoti-debug skill enables developers to debug models compiled with AOTInductor, providing visibility into the compilation process and runtime behavior. It facilitates troubleshooting by allowing inspection of generated code and execution traces within the PyTorch ecosystem.

When to use it

• You are encountering runtime errors in models optimized via AOTInductor and need to isolate the failure point.
• You want to inspect the intermediate representation or lowered code produced during the AOT compilation pipeline.
• Your team is debugging performance bottlenecks specific to compiled PyTorch graphs rather than eager mode execution.
• You need to validate that custom operators or kernels are correctly integrated into the AOTInductor build.

Key capabilities

• Debugging support for models compiled with AOTInductor.
• Inspection of compilation artifacts and generated code.
• Runtime tracing for optimized PyTorch workflows.

Example prompts

• "Run a debug session on this model compiled with AOTInductor to identify the source of the CUDA error."
• "Generate a trace of the execution flow for an AOT-compiled graph to check operator fusion logic."
• "Inspect the lowered code produced by AOTInductor for this custom layer to verify kernel registration."

Tips & gotchas

Ensure your PyTorch installation includes the latest torch.compile and AOTInductor features, as debugging tools depend on these being active. This skill is specifically designed for compiled workflows; it may not provide useful output when running models in standard eager mode without explicit compilation flags.