fix(z-image): dequantize GGUF pad tokens stored as a 2-D row

[Pfannkuchensack]

Summary

Z-Image GGUF transformers whose pad tokens are stored as a 2-D row (x_pad_token/cap_pad_token with GGUF shape [3840, 1], e.g. z-image-turbo-Q4_K_S.gguf) could not generate at all: every run failed in diffusers' _prepare_sequence with TypeError: Multiple dispatch failed for 'torch._ops.aten.where.self' … GGMLTensor. The GGUF key converter only dequantized pad tokens that were 1-D, so the 2-D ones stayed packed. Pad tokens of any single-row shape are now dequantized and reshaped to (1, dim); a pad token that is not a single row is refused with a clear error instead of being reshaped.

Related Issues / Discussions

Found while testing Z-Image ControlNets on a GGUF base.

QA Instructions

RTX 4090, z-image-turbo-Q4_K_S.gguf + Qwen3-4B + FLUX VAE, 8 steps: before, every generation failed with the error above (768×1344 and 1024², with and without ControlNet); after, both sizes generate normally.

Tests: a CPU end-to-end test writes a tiny GGUF with the released file's layout (Q8_0 layers, BF16 refiners/embedders, F32 norms, pad tokens F32 [dim, 1]), loads it through the real GGUF loader and compares the forward with a dense reference (fails with the same TypeError before the fix; zeroed or scrambled pad tokens also fail it); unit tests for 1-D, (dim,) and (1, dim) pad tokens and the refusal. pytest tests/backend/model_manager/load -k "z_image or gguf" and tests/backend/quantization/gguf: passed. Ruff clean.

Review

No material findings. Noted for follow-up, unchanged here: the Z-Image GGUF loader does not unpack small GGML tensors at load (other GGUF loaders do), and splitting the fused qkv keeps those weights dequantized in memory.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}