compile-overview

torch.compile pipeline works by progressing through Dynamo for Python capture, AOT Autograd for transformations, and Inductor for fusion and generated GPU or CPU code. Each stage produces distinct intermediate representations from Full ATen to Core ATen and Prims.

To debug graph breaks in torch.compile, use the compile-overview reference to select relevant TORCH_LOGS flags and inspect the appropriate stage output. This maps logging options to common investigations like graph breaks and poor fusion.

The torch.compile pipeline uses intermediate representations progressing from Full ATen to Core ATen, then to Prims, and finally to Inductor loop-level representations. This IR progression clarifies compilation behavior across Dynamo and Inductor stages.

Use stage-specific logging flags from TORCH_LOGS to investigate Inductor fusion issues. The compile-overview skill maps these flags, generated output files, and configuration options to specific compiler stages for selecting relevant logs.

Yes, you can use the compiler bisector workflows to troubleshoot PyTorch compilation performance issues. The skill provides bisector usage guidance alongside stage-specific logging flags and cache guidance for identifying compilation behavior problems.

torch.compile may generate poor fusion due to suboptimal Inductor transformations. Inspect the generated output files by mapping them to the appropriate stage using logging flags, enabling you to analyze the Inductor loop-level representations.