[Feature] Support video input for Kimi K2.5

[zyuhan1999]

Hi, Currently SGLang supports text and image inputs for Kimi K2.5, but not video. Could you please add support for video input inference for Kimi K2.5? Thanks!

[yhyang201]

I will support video input.

[wdx727]

I will support video input.

When do you plan to support it? Thanks!

[LuYanFCP]

Hello, have you been working on this feature recently? Due to work reasons, if you haven't arranged it, I would be happy to help implement this feature @yhyang201

[jiusanzhou]

🎉 Video input support for Kimi K2.5 is now ready!

I've implemented a comprehensive solution in PR #21023 that adds full video_url API support for Kimi K2.5.

✅ What's implemented:

• OpenAI-compatible video_url content type support
• Remote video URL downloads (HTTP/HTTPS)
• Multi-video and mixed content processing
• Full integration with Kimi K2.5's existing 3D vision tower
• Comprehensive error handling and fallbacks

📚 Usage Example:

{
  "model": "kimi-k2.5",
  "messages": [{
    "role": "user",
    "content": [
      {"type": "text", "text": "What's in this video?"},
      {"type": "video_url", "video_url": {"url": "https://example.com/video.mp4"}}
    ]
  }]
}

The implementation leverages SGLang's existing video download and processing infrastructure, making it robust and efficient.

📝 Status: PR is currently in draft for community review and testing. Once tested with real Kimi K2.5 deployments, it'll be ready for merge.

Please test and provide feedback! 🚀

[zyuhan1999]

🎉 Video input support for Kimi K2.5 is now ready!

I've implemented a comprehensive solution in PR #21023 that adds full video_url API support for Kimi K2.5.

✅ What's implemented:

• OpenAI-compatible video_url content type support
• Remote video URL downloads (HTTP/HTTPS)
• Multi-video and mixed content processing
• Full integration with Kimi K2.5's existing 3D vision tower
• Comprehensive error handling and fallbacks

📚 Usage Example:

{
"model": "kimi-k2.5",
"messages": [{
"role": "user",
"content": [
{"type": "text", "text": "What's in this video?"},
{"type": "video_url", "video_url": {"url": "https://example.com/video.mp4"}}
]
}]
}
The implementation leverages SGLang's existing video download and processing infrastructure, making it robust and efficient.

📝 Status: PR is currently in draft for community review and testing. Once tested with real Kimi K2.5 deployments, it'll be ready for merge.

Please test and provide feedback! 🚀

Thank you for your great work! May I ask how I can adjust the video_fps, max_video_frames, and image_size parameters in the online inference setting? In #21023, these parameters seem to be predefined.

[jiusanzhou]

Hi @zyuhan1999, thanks for the feedback!

The PR has been significantly refactored — we removed the custom video_processor.py and now reuse SGLang's standard multimodal pipeline (same approach as Qwen-VL and other models).

With the current implementation, video processing parameters are no longer hardcoded in our code. Instead, they are managed by:

1. HF Processor: The Kimi K2.5 HuggingFace processor handles video preprocessing (frame extraction, resizing, normalization) with its own default config from the model repository.

2. SGLang --mm-process-config: You can pass video processing config at server launch time via the --mm-process-config flag, e.g.:

   python -m sglang.launch_server \
     --model-path <model-path> \
     --mm-process-config '{"video": {"fps": 2, "max_frames": 8}}'

3. Per-request config: Some parameters can also be passed per-request in the video content part, following the same pattern as Qwen-VL.

The PR is currently in testing. Will update once validated with real Kimi K2.5 deployments.

[ygzhu]

🎉 Video input support for Kimi K2.5 is now ready!

I've implemented a comprehensive solution in PR #21023 that adds full video_url API support for Kimi K2.5.

✅ What's implemented:

• OpenAI-compatible video_url content type support
• Remote video URL downloads (HTTP/HTTPS)
• Multi-video and mixed content processing
• Full integration with Kimi K2.5's existing 3D vision tower
• Comprehensive error handling and fallbacks

📚 Usage Example:

{
"model": "kimi-k2.5",
"messages": [{
"role": "user",
"content": [
{"type": "text", "text": "What's in this video?"},
{"type": "video_url", "video_url": {"url": "https://example.com/video.mp4"}}
]
}]
}
The implementation leverages SGLang's existing video download and processing infrastructure, making it robust and efficient.

📝 Status: PR is currently in draft for community review and testing. Once tested with real Kimi K2.5 deployments, it'll be ready for merge.

Please test and provide feedback! 🚀

I deploy Kimi 2.5 according to the above PR and requested sglang data, but the model outputs answer as following:
I see that you've included a video placeholder (<|kimi_k25_video_placeholder|>), but I don't actually have access to the video content you're referring to. The placeholder appears to be a formatting token rather than viewable media.

To help you with a detailed description, I would need you to:

1. Upload the actual video file directly in this conversation (if supported by your interface)
2. Provide a link to the video (YouTube, Vimeo, etc.)
3. Describe what you're seeing or give me the title/context of the video so I can help analyze it.
   How should I fix this?