Woven by Toyota MLE Interview - PyTorch Debugging [D]

I have an upcoming interview with Woven by Toyota for an MLE perception role. I was told the round would be focused on ML coding/debugging. Specifically, I was told to focus on Python fundamentals, PyTorch, tensor operations and dimensions/shapes, common model-training code patterns, and debugging ML code.

Here are some common ways I've studied:

• Asking Claude to walk me through transformer, ViT, and CNN architectures (a basic one and U-Net) for my preparation. I would trace the shapes all the way through.
• Getting familiar with broadcasting rules (start from the left, see if the dimensions are equal, or are 1, and if so, take the greater one to get the resulting shape)
• Slicing tensors (i.e., how to get a column, how to get a row, etc.)
• Understanding why models can fail silently -- suspicious training vs val results, getting NaN for validation
• What a PyTorch loop looks like
• Common Python errors like indexing errors or value errors; iterators, generators, etc.

What are other suggestions? What architecture do they actually ask you about in this interview? Any help would be appreciated.