Forward pass of the model - how to pass videos?

I see this is how you run inference on the model:

```
output = model.generate(
input_ids,
return_dict_in_generate=True,
output_scores=True,
images_or_videos=videos,
modal_list=['video'],
max_new_tokens=128,
do_sample=True,
temperature=0.2,
use_cache=False,
)
```

I wanted to know how to run training code. The forward pass of the model does not support `modal_list` and `images_or_videos` - then how do I pass the videos? I want to be able to pass `labels` as I need to calculate probabilities. (so I need logits)

Start by reading the model's generate call and forward pass implementation, focusing on how images_or_videos, modal_list, labels, and logits are handled. Determine whether video inputs and label-based probability calculation are supported in the training path; done means a documented or working forward-pass path that accepts videos and produces logits for labels.

Written by the indexing model from the issue text.