Fix: Anima regional prompting should mirror width/height calculation for correct mask size

[FrankBro]

Summary

When using regional prompting with the Anima models, I would receive this error:

Server Error (2)
RuntimeError: The expanded size of the tensor (8160) must match the existing size (8040) at non-singleton dimension 2. Target sizes: [1, 16, 8160, 1024]. Tensor sizes: [1, 1, 8040, 1024]

An inverted ceil vs floor type division caused an off by 1 error.

QA Instructions

Claude used this script to confirm the math was wrong, based on my resolution (1080x1920) and the error message.

import math
for (w,h) in [(1080,1920),(1920,1080),(1024,1024)]:
    lh, lw = h//8, w//8
    old = (lh//2)*(lw//2)
    new = math.ceil(lh/2)*math.ceil(lw/2)
    def pad_tokens(d, p=2):
        pad = (p - d % p) % p
        return (d+pad)//p
    actual = pad_tokens(lh)*pad_tokens(lw)
    print(f'{w}x{h}: latent={lw}x{lh} old_mask={old} new_mask={new} transformer={actual} match={new==actual}')

# 1080x1920: latent=135x240 old_mask=8040 new_mask=8160 transformer=8160 match=True
# 1920x1080: latent=240x135 old_mask=8040 new_mask=8160 transformer=8160 match=True
# 1024x1024: latent=128x128 old_mask=4096 new_mask=4096 transformer=4096 match=True

I build my version of the code locally and can confirm the popup error does not appear anymore.
Additionally, I extracted the calculation into its own static function before adding a unit test to check for this regression.

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)

[lstein]

Reviewed the full diff and traced the fix through to consumption. This is a correct, well-scoped bugfix.

Correctness (verified): The new _compute_img_token_grid ceil division exactly mirrors the transformer's MiniTrainDIT._pad_to_patch_size, which pads the latent H/W up to a multiple of patch_spatial before patchifying. I traced the mask through preprocess_regional_prompt_mask — it resizes to (img_token_height, img_token_width) and flattens to img_seq_len, so mask and transformer now agree on the ceil-padded token count. This is the only token-grid computation on the regional path, so nothing else needed the same fix.

Tests: test_matches_transformer_padding is the highlight — it cross-checks the computed grid against the transformer's real padding logic (via a lightweight SimpleNamespace stub) rather than a hand-copied expectation, so the two sides can't silently drift. The explicit 8160-token regression case documents the original crash nicely. All 24 tests pass locally.

One small follow-up I pushed directly (1e9ff93, maintainerCanModify was on): two docstrings — in _load_text_conditionings and preprocess_regional_prompt_mask — still described the grid as H // patch_size (floor), which is exactly the thing this PR fixes. Updated them to ceil(latent / patch_size) so the docs match the math. No functional change.

Nice work extracting the computation into a testable static method — that's the right shape for this fix. LGTM.