Fresh runs can save a checkpoint at step 0 and tensorboard writes unconditionally

[janmarxen]

Bug report

MaxText appears to perform checkpoints even when setting flags to disable it. In my case, the run used load_parameters_path, enable_checkpointing=True (to enable initial parameter loading), checkpoint_period=1000000 save_checkpoint_on_completion=False, and still attempted checkpoint-related work in a 50 step training run.

The same issue applies to enable_tensorboard, which is ignored by write_setup_info_to_tensorboard.

@giusgal and @janmarxen

Logs/Output

No response

Environment Information

No response

Additional Context

No response

[xuefgu]

If you would like to skip saving checkpoints entirely, you can specify enable_checkpointing=False. It doesn't affect using load_parameters_path to load a checkpoint. So that should address the first concern.

The tensorboard concern should have been addressed by #3806.