Training freezes when trying to evaluate previously saved checkpoint

I'm fine-tuning an mT5 small checkpoint on a custom mixture. For debugging, I've set `eval_period=25` and `utils.SaveCheckpointConfig.period=50` in finetune.gin.
The training job freezes inside `trainer.py` whenever it tries to compute metrics on a saved checkpoint (see attached screenshot)


Start with finetune.gin and the SaveCheckpointConfig.period setting, then inspect trainer.py where metrics are computed for saved checkpoints. Reproduce the freeze using eval_period=25 and checkpoint period=50, and trace that evaluation path. Done means training completes checkpoint evaluation and reports metrics instead of freezing.

Written by the indexing model from the issue text.