Handle prompt embedding concat in Qwen dreambooth example

[chenyangzhu1]

What does this PR do?

When running the train_dreambooth_lora_qwen_image.py script with the --with_prior_preservation flag, the training crashes during the text embedding extraction phase. Fixing two distinct bugs related to tensor concatenation at L1323 and L1324.

Solution

To resolve these issues, the tensor concatenation logic needs to handle dynamic padding and missing masks dynamically:

1. Reconstruct Masks: If a mask is missing (None), reconstruct it as an all-ones tensor based on its corresponding embedding's sequence length.
2. Pad to Max Length: If the sequence lengths differ, pad the shorter tensor(s) to match the longest sequence.
3. Fold Back to None: After concatenation, if the merged mask consists entirely of 1s, fold it back to None.

Fixes #13386

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case. [Bug] train_dreambooth_lora_qwen_image.py crashes with --with_prior_preservation due to tensor concatenation errors #13386
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@sayakpaul

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[sayakpaul]

Would maybe also add a comment on how this is only relevant during prior preservation.

[chenyangzhu1]

Done.

I added the following comment

# These helpers only matter for prior preservation, where instance and class prompt
# embedding batches are concatenated and may not share the same mask/sequence length.

[sayakpaul]

Thanks! Just one comment.

[sayakpaul]

@bot /style

[github-actions]

Style bot fixed some files and pushed the changes.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.