Hints

DL-C16-C-001

Stable softmax subtracts the maximum.

DL-C16-C-002

Subtract 10 from the first logit.

DL-C16-C-003

Add the two values.

DL-C16-C-004

Large exponentials are dangerous.

DL-C16-C-005

NaN means "not a number."

DL-C16-C-006

Think of storage format and representable values.

DL-C16-C-007

Mixed precision is a controlled tradeoff.