Hyperparameter sweep "crashed" logging is false

[kylebgorman]

If a hyperparameter sweep run uses patience to stop, you'll see something like the following logged:

sweep.py INFO: 27-Dec-24 01:15:00 - Monitored metric val_loss did not improve in the last 10 records. Best score: 0.543. Signaling Trainer to stop.

However, W&B treats this is as a "crash" rather than a successful halt. This is strange. Perhaps when patience halts training, it returns a non-zero error code. Assuming this is a unique error code and different than the error code used for actual crashes, we could intercept it here to eliminate this.

[kylebgorman]

My hypothesis about what is happening is wrong. Early stopping doesn't give you a non-zero error code.