Follow-up: clean up video_duration handling in hyperclovax_vision_v2 (#44314)

[jp1924]

Follow-up from #44314.

HyperCLOVAXVisionV2Processor currently fills video_duration with a processor-side stopgap (video_duration_token + overrides on apply_chat_template/_process_videos in modular_hyperclovax_vision_v2.py) instead of the single-placeholder + replace_video_token pattern other video models use (qwen3_vl, glm4v).

Why it's a stopgap and not the real fix (from the PR discussion, #44314 (comment)):

• vLLM: the single-placeholder pattern pulls video_metadata through mm_processor_kwargs, and there have been cases of metadata/frame-count sync issues on the vLLM side (e.g. Qwen3-VL). Moving the structure into replace_video_token would need a matching vLLM-side change.
• Consistency: the hub chat_template.jinja currently renders image and video symmetrically (both emit a mime block + PAD via jinja). Collapsing only video to a single token breaks that symmetry; fixing it properly means reworking the image path too, which was out of scope for add HyperClovaX Vision #44314.

Real fix path, roughly in order:

1. hub chat_template.jinja change (Naver side)
2. vLLM compatibility check, since the template is shared with vLLM
3. transformers-side migration to replace_video_token, dropping the stopgap in modular_hyperclovax_vision_v2.py

cc @bigshanedogg @vasqu

[Kaustubh1235]

I will take this issue. Please assign it to me.

The current handling of video_duration in HyperCLOVAXVisionV2Processor uses a stopgap approach. I will first review the modular_hyperclovax_vision_v2.py file to understand how video_duration is being processed. Then, I will look into aligning it with the replace_video_token pattern used in other models like qwen3_vl. The goal is to ensure consistency across models while maintaining functionality. I will verify the changes by running tests to confirm that video metadata is correctly processed.