set timestep index to 0 to avoid lookup in scheduler.step

[bghira]

This pull request addresses a potential issue with floating-point precision in the scheduler's timestep lookup across multiple model pipelines. The main change is the addition of a call to self.scheduler.set_begin_index(0) before entering the denoising loop in each pipeline. This ensures robust and consistent behavior during inference by preventing failures in scheduler.step() due to precision errors, as referenced in a related HuggingFace diffusers pull request.

Key changes by theme:

Scheduler Robustness and Consistency:

• Added self.scheduler.set_begin_index(0) before the denoising loop in the __call__ or equivalent method of the following pipelines to avoid floating-point precision issues in timestep lookup (see: Avoid DtoH sync from access of nonzero() item in scheduler huggingface/diffusers#11696):
  • auraflow/pipeline.py
  • auraflow/pipeline_controlnet.py
  • chroma/pipeline.py
  • cosmos/pipeline.py
  • flux/pipeline.py [1] [2]
  • flux/pipeline_controlnet.py [1] [2]
  • hidream/pipeline.py
  • hunyuanvideo/pipeline.py
  • hunyuanvideo/pipeline_i2v.py
  • kandinsky5_image/pipeline_kandinsky5_t2i.py
  • kandinsky5_video/pipeline_kandinsky5_t2v.py
  • longcat_image/pipeline.py
  • longcat_image/pipeline_edit.py
  • longcat_video/pipeline.py
  • ltxvideo2/pipeline_ltx2.py
  • ltxvideo2/pipeline_ltx2_image2video.py
  • wan/pipeline.py [1] [2]
  • wan_s2v/pipeline.py
  • z_image/pipeline.py
  • z_image_omni/pipeline.py