Fix Flux2 DreamBooth prior preservation prompt repeats

[azolotenkov]

What does this PR do?

Fixes a prior-preservation batch size mismatch in the Flux2 DreamBooth LoRA scripts.

When custom instance prompts are not used, prompt_embeds already contains concatenated instance + class embeddings under --with_prior_preservation, but collate_fn also doubles prompts, so repeating by len(prompts) over-expands text embeddings by 2x and no longer matches the latent batch size.

This applies the same fix pattern as #13396 to:

• train_dreambooth_lora_flux2.py
• train_dreambooth_lora_flux2_klein.py

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.
• 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

[Copilot]

Pull request overview

Fixes a prior-preservation batch size mismatch in the Flux2 DreamBooth LoRA training example scripts by adjusting how many times static text embeddings are repeated when --with_prior_preservation is enabled.

Changes:

• Adjust num_repeat_elements to use len(prompts) // 2 under --with_prior_preservation (since collate_fn doubles prompts).
• Apply the same fix to both Flux2 DreamBooth LoRA variants (standard + klein).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
examples/dreambooth/train_dreambooth_lora_flux2.py	Fixes prompt-embed repeat count for prior preservation when using static (non-custom) prompts.
examples/dreambooth/train_dreambooth_lora_flux2_klein.py	Same repeat-count fix in the klein variant.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[azolotenkov]

Updated again.

The previous repeat_interleave(..., dim=0) change fixed the prompt/text ordering issue for --with_prior_preservation with train_batch_size > 1, but while testing it locally I hit another mismatch in the same code path: weighting was still left at full batch size after model_pred and target were chunked into instance/prior halves.

This update chunks weighting alongside model_pred and target in both Flux2 scripts.

Validation used:

• hf-internal-testing/tiny-flux2-klein
• --with_prior_preservation
• --train_batch_size 2

That setting now completes successfully for klein, and it exercises exactly the batch-layout path discussed in Copilot review.

[azolotenkov]

@sayakpaul, this PR is the same fix pattern that was now merged for Qwen Image DreamBooth in #13441.

[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.

[sayakpaul]

Failing tests are unrelated.