Feature request – expose full video‑input support & raise 5‑image limit (Qwen 2.5‑VL & InternVL3)

[Martin-Hausleitner]

Description:
Both Qwen 2.5‑VL and InternVL3‑78B natively support direct video ingestion and fine‑grained frame control.
Official docs & examples:

• Qwen2.5‑VL Quickstart (video message with URL, fps, num_frames, initial_shift, pixel budgets)
  https://github.com/QwenLM/Qwen2.5-VL#quickstart
• InternVL3 video sampling (num_frames, fps, initial_shift)
  https://huggingface.co/docs/transformers/en/model_doc/internvl

Parameters that should be configurable via Chutes API

Category	Parameter(s)	Doc reference
Frame sampling	fps, num_frames, initial_shift	Qwen README, InternVL docs
Resolution caps	min_pixels, max_pixels, total_pixels	Qwen README
Resizing	resized_width, resized_height (multiples of 28)	Qwen README
Backend	FORCE_QWENVL_VIDEO_READER = torchvision|decord|torchcodec	Qwen repo code
Token IDs	image_token_id, video_token_id	Qwen README

Example request body (desired)

{
  "model": "Qwen/Qwen2.5-VL-72B-Instruct",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "video",
          "video_url": "https://example.com/video.mp4",
          "fps": 2,
          "num_frames": 8,
          "initial_shift": true,
          "total_pixels": 20480,
          "min_pixels": 784,
          "max_pixels": 100352,
          "resized_width": 448,
          "resized_height": 448,
          "FORCE_QWENVL_VIDEO_READER": "decord"
        },
        { "type": "text", "text": "Describe the video." }
      ]
    }
  ]
}