Fixing issue with video loading for Gemma 4 with relative paths

[perone]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

When using datasets with relative media paths and --root_image_dir, video and audio file paths are not resolved to absolute paths before being passed to model-specific templates. Image paths are correctly resolved via _load_image, but video and audio paths are passed as raw strings.

This causes a FileNotFoundError for models like Gemma 4 whose templates delegate media loading to the HuggingFace processor, since the processor receives the unresolved relative path.

This fix adds path resolution for inputs.videos and inputs.audios in Template._preprocess_media (swift/template/base.py), matching the existing behavior for images. A relative path is resolved against root_image_dir only when:

• root_image_dir is set
• The path is a string (not already loaded)
• The file does not exist at the original path
• The path is not a URL
• The resolved path points to an existing file

Experiment results

Verified that Gemma 4 video fine-tuning with relative video paths in the dataset and --root_image_dir correctly loads videos after this fix, where previously it raised FileNotFoundError.

[gemini-code-assist]

Code Review

This pull request adds logic to resolve video and audio file paths using ROOT_IMAGE_DIR to ensure absolute paths are available for model-specific templates. The reviewer suggested refactoring the code to eliminate duplication between the video and audio path resolution loops.

[Jintao-Huang]

thanks!

[Jintao-Huang]

There is also the case where the value is base64-encoded. Could you fix this by referencing the handling here:

ms-swift/swift/template/vision_utils.py

Line 104 in 514ce7c

def _check_path(path: str) -> Union[str, None]:

[Jintao-Huang]

Never mind, I'll open a PR to fix it.

[perone]

Never mind, I'll open a PR to fix it.

Sorry just saw it now, thanks again !