Problem

startup_time is calculated (line 613) but never printed in the final summary. This information is useful for:

• Understanding torch.compile overhead
• Debugging slow startup issues
• Calculating total experiment time vs training time

Solution

Add startup_seconds to the final summary output.

Example Output

---
val_bpb:          0.997900
training_seconds: 300.1
total_seconds:    325.9
...
depth:            8
startup_seconds:  25.8