Wan2.1-VACE-14B Inference Bug

Thank you for your excellent work, but when running Wan2.1-VACE-14B inference, I encountered the following error when setting vace_video, vace_video_mask, and vace_reference_image simultaneously.

Traceback (most recent call last):
  File "/DiffSynth-Studio-New/examples/wanvideo/model_inference/Wan2.1-VACE-14B.py", line 42, in <module>
    video = pipe(
  File "/mnt/bn/maliva-gen-ai-v2/liwei.947/software/anaconda/envs/wanx/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
    return func(*args, **kwargs)
  File "/DiffSynth-Studio/diffsynth/pipelines/wan_video_new.py", line 449, in __call__
    noise_pred_posi = self.model_fn(**models, **inputs_shared, **inputs_posi, timestep=timestep)
  File "/DiffSynth-Studio/diffsynth/pipelines/wan_video_new.py", line 978, in model_fn_wan_video
    vace_hints = vace(x, vace_context, context, t_mod, freqs)
  File "/mnt/bn/maliva-gen-ai-v2/liwei.947/software/anaconda/envs/wanx/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/mnt/bn/maliva-gen-ai-v2/liwei.947/software/anaconda/envs/wanx/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
    return forward_call(*args, **kwargs)
  File "/DiffSynth-Studio/diffsynth/models/wan_video_vace.py", line 60, in forward
    c = torch.cat([
  File "/DiffSynth-Studio/diffsynth/models/wan_video_vace.py", line 61, in <listcomp>
    torch.cat([u, u.new_zeros(1, x.shape[1] - u.size(1), u.size(2))],
RuntimeError: Trying to create tensor with negative dimension -7200: [1, -7200, 5120]

I've noticed others experiencing the same error in Issues 1 and Issues 2. The resolution of my three inputs is consistent, and the total frame count of both videos is the same, so it seems unrelated to the inputs. I also tried the official VACE code, which worked fine.

I'm looking forward to your response and thank you for your help.

Start with examples/wanvideo/model_inference/Wan2.1-VACE-14B.py and diffsynth/models/wan_video_vace.py, then trace the call through diffsynth/pipelines/wan_video_new.py:978 using the three simultaneous inputs. Reproduce the negative-dimension failure and compare input-shape handling with the official VACE implementation; done means Wan2.1-VACE-14B inference succeeds for that combination without the runtime error.