fix: stop default collater from adding padding_mask

[zeel2104]

What does this PR do ?

Stop default_collater from creating a padding_mask for generic padded batches, so models that do not accept that argument no longer need downstream filtering and do not crash.

Changelog

• Remove synthesized padding_mask creation from nemo_automodel.components.datasets.utils.default_collater
• Update the default collator unit test to validate the new batch contract
• Keep the fix scoped to the generic collator path without changing THD/context-parallel handling

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to workaround required for padding_mask #1639

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[hemildesai]

Hi, thanks a lot for your contribution.

We need padding mask in the batch for correctly calculating aux_loss for MoEs. Rn, padding mask needs to be passed explicitly to handle some corner cases. As a result, I think an appropriate fix would be to add padding_mask to the batch if the model supports it, skip it otherwise.

[zeel2104]

@hemildesai
Thanks for the clarification. I’ve updated the fix accordingly.

default_collater still produces padding_mask, but the dataloader path now only keeps/passes it when the target model supports padding_mask in forward(). For models that do not support it, the batch drops padding_mask before the forward path.

This keeps the MoE aux-loss path intact while avoiding crashes for models that cannot consume the argument.

I also added targeted coverage for:

• detecting whether a model supports padding_mask
• keeping padding_mask for supporting models
• stripping it for unsupported models
  Let me know if anything more needs to be done

[akoumpa]

@adil-a + @hemildesai can you provide guidance? Thank you.