Make HunyuanDiT2DModel and HunyuanDiT2DControlNetModel hidden states contiguous

[menglcai]

What does this PR do?

Similar to #14186

The non-contiguous tensor originates from PatchEmbed.forward — flatten(2).transpose(1, 2) (BCHW → BNC) produces a non-contiguous layout whose overhead accumulates across transformer blocks in HunyuanDiT2DModel and HunyuanDiT2DControlNetModel.
Unlike SD3 where the benefit is limited to ROCm, this patch improves performance on both ROCm and NVIDIA for HunyuanDiT, though the gain on NVIDIA is smaller.

Benchmarks (HunyuanDiT-v1.1-Diffusers, 512×512, 50 steps, fp16):

Platform	PyTorch	Default	+contiguous	Speedup
NVIDIA RTX 5090	2.12.1+cu132	3.16 it/s	3.29 it/s	+4.1%
AMD Radeon AI PRO R9700 (gfx1201)	2.12.0+rocm7.15.0a20260713	1.34 it/s	1.52 it/s	+13.4%
AMD Radeon(TM) 8060S Graphics (gfx1151)	2.12.0+rocm7.15.0a20260711	3.23 s/it	3.05 s/it	+5.8%

Benchmarks (HunyuanDiT-v1.2-Diffusers, 512×512, 50 steps, fp16):

Platform	PyTorch	Default	+contiguous	Speedup
NVIDIA RTX 5090	2.12.1+cu132	3.16 it/s	3.30 it/s	+4.4%
AMD Radeon AI PRO R9700 (gfx1201)	2.12.0+rocm7.15.0a20260713	1.35 it/s	1.53 it/s	+13.3%
AMD Radeon(TM) 8060S Graphics (gfx1151)	2.12.0+rocm7.15.0a20260711	3.23 s/it	3.05 s/it	+5.6%

Benchmarks (HunyuanDiT-v1.2-ControlNet-Diffusers-Canny + HunyuanDiT-v1.2-Diffusers-Distilled, 512×512, 50 steps, fp16):

Platform	PyTorch	Default	+contiguous	Speedup
NVIDIA RTX 5090	2.12.1+cu132	2.12 it/s	2.26 it/s	+6.6%
AMD Radeon AI PRO R9700 (gfx1201)	2.12.0+rocm7.15.0a20260713	1.13 s/it	1.04 it/s	+8.0%
AMD Radeon(TM) 8060S Graphics (gfx1151)	2.12.0+rocm7.15.0a20260711	4.73 s/it	3.79 s/it	+19.9%

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• 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?
• Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

@sayakpaul @yiyixuxu

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]

Thanks for your PR. Instead of Hunyuan, do we want to treat another model that has popularity and more usage?

[github-actions]

Hi @menglcai, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder.

[menglcai]

Thanks for your PR. Instead of Hunyuan, do we want to treat another model that has popularity and more usage?

Understood, I'll look into more popular or widely used models and open a follow-up PR if a similar fix applies.

[menglcai]

Hi @sayakpaul , I've found similar patterns (some tensors not contiguous before the loop) with potential for improvement (still needs careful testing):

• SD3ControlNetModel (for stabilityai/stable-diffusion-3-medium-diffusers) — roughly a 2–3% improvement for ROCm pytorch
• WanTransformer3DModel (for Wan-AI/Wan2.2-TI2V-5B-Diffusers) — roughly a 10–15% improvement for ROCm pytorch

For newer models such as Qwen-image, FLUX.2-klein-4B, ERNIE-Image, and Z-Image-Turbo, this pattern was not observed.

Would it make sense to continue pushing this PR (HunyuanDiT2DModel), or would you suggest closing it for now and focusing on the other models first? Thank you

[sayakpaul]

would you suggest closing it for now and focusing on the other models first? Thank you

Because this would be more impactful and thanks so much for your willingness!