[MultiModal] Current Megatron-Bridge Does Not Support Video Modal

[Zhuohao-Li]

Description

Current docker built from https://github.com/fzyzcjy/Megatron-Bridge/tree/dev_rl (0.3.0.rc0) does not support video pad input, which occurs:

  File "/usr/local/lib/python3.12/dist-packages/megatron/bridge/models/qwen_vl/modelling_qwen3_vl/model.py", line 211, in forward
    assert pixel_values_videos is None and video_grid_thw is None, "not support video now"

Solution

Upgrade to latest 0.4.0.rc0 fix this. Will solve this by a PR later

[coding-famer]

And a note on future video input support: sglang has a resize&resample preprocessing step for video input (Qwen-VL), and we need to ensure the video inputs are consistent between training and inference.