[pthombre]

• Add text-to-video multiresolution dataloader: Introduces TextToVideoDataset and build_video_multiresolution_dataloader to support video training (Wan, HunyuanVideo) with bucket-based multiresolution sampling, replacing the legacy build_dataloader / MetaFilesDataset path.
• Refactor diffusion datasets with shared base class: Extracts common multiresolution logic (metadata loading, bucket grouping, dynamic batch sizing) into BaseMultiresolutionDataset, with TextToImageDataset and TextToVideoDataset as concrete implementations.
• Refactor and rename dataloader APIs: Renames collate_fn_flux → collate_fn_text_to_image, build_flux_multiresolution_dataloader → build_text_to_image_multiresolution_dataloader, and extracts a shared _build_multiresolution_dataloader_core helper. Moves collate_fn_production and dataloader builders from sampler.py
  to collate_fns.py.
• Add video collate function (collate_fn_video) with support for model-specific optional fields (e.g. text_mask, image_embeds).
• Update example configs: Migrate Wan 2.1 and HunyuanVideo YAML configs to use the new video multiresolution dataloader; update Flux configs to use renamed image dataloader API.
• Remove obsolete CI/CD nightly scripts and configs for Wan 2.1 pretrain.
• Fix Flux attention backend configuration (use "flash" instead of "_flash_3_hub").
• Fix and update unit tests for renamed APIs and new video dataloader.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[pthombre]

/ok to test a31c392

[pthombre]

/ok to test fa92b0c

[pthombre]

/ok to test 755c134

[pthombre]

/ok to test 4a891b8