Here are the first five measurements to take when investigating generation slowdown at long context:

1. **Time to first token (