Unstable loss after a learning-rate increase

After increasing the learning rate, training loss oscillates and occasionally becomes non-finite. What would you investigate first?

Compare against the previous rate on identical data order and initialisation. Inspect gradient norms, input scales, loss implementation and mixed-precision behaviour, and check for non-finite inputs. A rate that overshoots can cause instability, but data or numerical errors can look similar. Reduce the rate, consider justified gradient clipping, and verify a small deterministic example before expanding the search.