train qwen3-vl-moe on ShareGPT4V-small with quick-start

[iqiancheng]

Summary

This PR adds support for resolving relative image/video/audio paths relative to the train_path directory, making it easier to use simple datasets like ShareGPT4V-small without requiring absolute paths in data files.

Changes

Bug Fixes

• Fix AttributeError: 'DataArguments' object has no attribute 'mm_configs' in train_qwen_vl.py by using MyDataArguments instead of DataArguments in the Arguments class

New Features

• Add resolve_relative_path() utility function in veomni/data/multimodal/file_utils.py to handle unified relative path resolution
• Update image_utils.py, video_utils.py, and audio_utils.py to use the new path resolver
• Pass train_path through kwargs in train_qwen_vl.py to enable relative path resolution for Qwen2.5-VL and Qwen3-VL models

Documentation

• Add training guide docs/examples/qwen3vl_moe.mdlink for Qwen3-VL MoE model

Behavior

When image/video/audio paths in dataset files are relative (e.g., coco/train2017/image.jpg), they are automatically resolved relative to the directory containing train_path. This allows datasets like ShareGPT4V-small to work without requiring absolute paths.

Example:

• train_path: /path/to/ShareGPT4V-small-coco-128.jsonl
• Image path in dataset: coco/train2017/image.jpg
• Resolved path: /path/to/coco/train2017/image.jpg

Testing

Tested with ShareGPT4V-small dataset using relative image paths.

Related Issues

Fixes issues related to multimodal data file path resolution for easier dataset setup. #186 #191

[CLAassistant]

All committers have signed the CLA.

[Luosuu]

Thank you for your contribution! I will request an additional reviewer