compile-trace-aot

To debug incorrect gradients in torch.compile training, inspect the AOT Autograd joint and partitioned FX graphs. This Skill traces functionalization and forward-backward graph generation to validate mutation handling and gradient flow against eager execution.

AOT Autograd functionalization transforms mutation operations into functional forms for joint graph generation. You need to trace it when compiled training produces incorrect gradients or unexpected memory usage, requiring visibility into saved activations and partitioning correctness.

Inspect PyTorch FX graphs by tracing AOT Autograd logging outputs to verify saved activations and recomputation behavior. This Skill makes post-gradient optimization passes visible, helping you compare eager and compiled memory usage to identify partitioning issues.

Yes, you can trace graph partitioning and recomputation behavior in PyTorch AOT Autograd. This Skill interprets PyTorch logging to validate how joint forward and backward graphs are split and how recomputation affects gradient correctness and memory usage.

torch.compile produces different gradient behavior when AOT Autograd transformations alter mutation handling or saved activations during functionalization. This Skill helps compare eager and compiled FX graphs to pinpoint discrepancies in joint forward and backward graph partitioning.

Tracing PyTorch AOT Autograd transformations requires interpreting detailed FX graph outputs and PyTorch logging. It is limited to validating functionalization, partitioning, and post-gradient optimization, meaning it focuses on graph-level debugging rather than underlying tensor operations.