fix: fix the bug of online serving

[Benedict-Y]

PR: Fix TypeError in Online Serving + vLLM video inference parameters

Hi Kwai Team, thank you for the great work and for reviewing this PR.

• Related issue: Bug: TypeError: process_vision_info() got an unexpected keyword argument 'return_video_kwargs' #42

Summary

• Fix a TypeError in the Online Serving demo caused by passing an unsupported return_video_kwargs argument to process_vision_info.
• Add recommended vLLM serve flags for stable video inference.

Changes

• Remove return_video_kwargs=True from the demo/serving code.
• Keep the call as:

  image_inputs, video_inputs, video_kwargs = process_vision_info(video_message)

  The function already returns video_kwargs.

Rationale

• process_vision_info in src/keye_vl_utils/vision_process.py doesn’t accept return_video_kwargs; passing it raises:

  TypeError: process_vision_info() got an unexpected keyword argument 'return_video_kwargs'
  

vLLM Notes (to avoid freezes with long videos)

vllm serve Kwai-Keye/Keye-VL-8B-Preview \
  --tensor-parallel-size 1 \
  --enable-prefix-caching \
  --gpu-memory-utilization 0.6 \
  --host 0.0.0.0 \
  --port 8000 \
  --max-num-batched-tokens 80960 \
  --max-model-len 80960 \
  --trust-remote-code

• --max-num-batched-tokens 80960: avoids batching issues with long sequences
• --max-model-len 80960: supports extended context from video frames

Testing

• Re-ran the Online Serving demo with a sample video.
• Verified:
  • No TypeError occurs.
  • Chat completion succeeds.