internvl inference: how set max_num_frames value?

[Bendywater]

When I inference video based internvl model, I find the frames of video is not transformer in vllm and only there are 32 frames. Therefore, I want to know how to set num_frames for video. I try this server: --media-io-kwargs '{"video": {"num_frames": 60}}' but I don't konw if it is right.

env CUDA_VISIBLE_DEVICES=2 vllm serve /mnt/models/InternVL3_5-4B \
    -dp 1 \
    --gpu-memory-utilization 0.8 \
    --max-model-len 40K \
    --served-model-name InternVL3_5-4B \
    --media-io-kwargs '{"video": {"num_frames": 60}}' \
    --port 12400 \
    --trust-remote-code
``

[Tahirc1]

@Bendywater

i tried running model with given command and with different num frames output was not different i ran inference on video of 9 seconds @ 30 fps and it gave proper summary

 vllm serve OpenGVLab/InternVL3_5-4B \
  --host 0.0.0.0 \
  --max-model-len 32768 \
  --limit-mm-per-prompt ' {"video": {"count": 1, "num_frames": 30, "width": 512, "height": 512}, "image": {"count": 5, "width": 512, "height": 512}}' \
  --trust-remote-code

--

vllm serve OpenGVLab/InternVL3_5-4B \
  --host 0.0.0.0 \
  --max-model-len 32768 \
  --media-io-kwargs '{"video": {"num_frames": 60}}' \
  --trust-remote-code

--

You can checkout the documentation at

https://docs.vllm.ai/en/latest/api/vllm/config/multimodal/#vllm.config.multimodal.MultiModalConfig

https://docs.vllm.ai/en/latest/cli/serve/#-limit-mm-per-prompt

https://docs.vllm.ai/en/latest/cli/serve/#-media-io-kwargs