fix(z-image): handle missing unified mask in regional guidance forward

[Zhuoxi2000]

Summary

Bug fix. Z-Image Turbo generations with a Regional Guidance layer fail with AttributeError: 'NoneType' object has no attribute 'bool' at invokeai/backend/z_image/z_image_transformer_patch.py:160.

Why: #9550 bumped diffusers from 0.39 to 0.40. In 0.40, ZImageTransformer2DModel._build_unified_sequence returns attn_mask=None when every batch item has the same unified length:

if all(seq == max_seqlen for seq in unified_seqlens):
    attn_mask = None

z_image_denoise always calls the transformer with a single item (cap_feats=[pos_prompt_embeds]), so inside the patched regional forward unified_mask is always None and unified_mask.bool() raises on every regional guidance run. With 0.39 the single-item mask was all True.

How: when unified_mask is None, build the base padding mask as all-valid (torch.ones((bsz, unified_seqlen), dtype=torch.bool)), which is what None means in diffusers and matches the 0.39 behaviour. The odd layers already pass unified_mask (None) straight to the layers, so only the float-mask construction needed the change. +8/-1 in the patch file.

Related Issues / Discussions

Closes #9612

QA Instructions

Unit tests (CPU, tiny randomly initialised ZImageTransformer2DModel, no weights needed):

pytest -q tests/backend/z_image/test_z_image_transformer_patch.py

• test_regional_forward_single_item_batch: single item, 16 image + 5 text tokens, regional mask that blocks part of the image from the prompt. Regression test for [bug]: Z-Image turbo regional guidance. #9612.
• test_regional_forward_single_item_batch_matches_unpatched_without_padding: 64 image + 32 text tokens (no SEQ_MULTI_OF padding). An all-True regional mask gives the same output as the unpatched diffusers forward, and a restrictive mask changes the output, so the regional mask still takes effect.

Results with diffusers 0.40.0 (the pinned version):

• on main (a51e26f): 2 failed, both with AttributeError: 'NoneType' object has no attribute 'bool' at z_image_transformer_patch.py:160
• with this PR: 3 passed for tests/backend/z_image/
• ruff check and ruff format --check pass on both files

Manual check for reviewers (I have only run the unit tests above, not a full generation with Z-Image Turbo weights): Canvas, Z-Image Turbo model, add one Regional Guidance layer with a prompt and a painted region, Invoke. Before this PR the generation fails with the error above; with it, generation should complete with the regional prompt applied to the painted area.

Merge Plan

N/A

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

AI assistance: this change was drafted with an AI coding assistant (Claude) and verified locally with the tests above.

[Zhuoxi2000]

Closing: this is already fixed on main by d6a54b6 ("fix(z-image): regional prompting at batch size 1"), which landed with #9613. Thanks!