Catch ConfigError when attempting to set num_parameters in WANDB

[gheinrich]

What does this PR do?

The current version of the code catches AttributeError when attempting to set the model's number of parameters in the Weights & Biases config.

ConfigError is another exception that needs catching in the case when the current number of model parameters differs from that of the config. This happens, for example, on resuming from checkpoints if extra layers have been added.

Example error message:

wandb.sdk.lib.config_util.ConfigError: Attempted to change value of key "model/num_parameters" from 0 to 700416

Since the current code already includes support for ignoring AttributeError, it feels safe to add this extra catch.

I did not find a test suite for the WANDB integration.

[gheinrich]

Hello, cc @parambharat since you wrote the code that catches the first exception. Thank you!

[LysandreJik]

Is there a path for the ConfigError that doesn't depend on self?

[vasqu]

Similar to #33464, but we could import the error at __init__ instead imo:

transformers/src/transformers/integrations/integration_utils.py

Lines 764 to 769 in d667888

	def __init__(self):
	has_wandb = is_wandb_available()
	if not has_wandb:
	raise RuntimeError("WandbCallback requires wandb to be installed. Run `pip install wandb`.")
	if has_wandb:
	import wandb

[gheinrich]

Hello thanks for the comment, yes I tried to remain consistent with the other wandb references in this file which are all made through self._wandb. Feel free to close this PR in favor of #33464 if you prefer the other implementation :-)

[LysandreJik]

Thanks both! IMO it's a bit cleaner to have the error be imported here rather than as a reference to self, so I'll go ahead and merge the other PR.

I appreciate your contribution, please keep them coming!

[gheinrich]

Thank you, I will abandon this PR.