Debug a Transformer bug in an unfamiliar repo

You are given an unfamiliar GitHub repository that implements a Transformer model. The interviewer claims there is a bug causing one of the following symptoms:

• training loss diverges or becomes NaN,
• model quality is far worse than expected,
• outputs violate causality for an autoregressive model,
• shapes/broadcasting silently behave incorrectly.

You cannot rewrite the whole project; your task is to find and fix the bug efficiently.

Explain:

1. The step-by-step debugging strategy you would use.
2. What minimal experiments/tests you would run to localize the issue.
3. The most common Transformer implementation bugs you would check first.
4. How you would validate the fix and prevent regressions.