What problem does it solve? Training runs fail in opaque ways: NaN losses, missing gradients, misleading loss values from shape mismatches, and non-reproducible results. This Skill turns a reported symptom into an observable, bounded failure so the root cause can be identified and verified instead of masked. ## Core Features & Use Cases - Failure isolation: Reduce a failing training run to a fixed batch while preserving shapes, label semantics, and value ranges, then trace finite values through inputs, loss, gradients, and parameter updates. - Framework-specific controls: Apply the correct debugging levers for PyTorch (autograd, anomaly detection, AMP), Lightning (fast_dev_run, manual vs automatic optimization), and TensorFlow/Keras (GradientTape, run_eagerly, op determinism). - Verified fixes: Compare baseline and candidate fix from equivalent model and optimizer state with unchanged assertions and principled numerical tolerances. - Use Case: Your PyTorch training loss looks wrong despite correct predictions. The Skill identifies a broadcasting bug where [N, 1] predictions minus [N] labels produce an [N, N] residual, then verifies the shape fix against the baseline. ## Quick Start Debug my PyTorch training run where the loss suddenly becomes NaN after a few steps and tell me the root cause with a verified fix.
