Wrong aspect_ratio param in llava-onevision evaluation

[Luodian]

I think the param is not set correctly.

MEGA-Bench/megabench/models/LlavaOV.py

Line 68 in ee2e9f3

overwrite_config["image_aspect_ratio"] = "pad"

It should be "anyrex_max_9" or directly read from the config.json

[woodfrog]

We followed your official tutorial for multi-image interleaved input. See the section of Image-Text Interleaved Input.

Do you have thoughts/insights on how this parameter setting would affect the performance? We can run again with anyrex_max_9 if the one in the tutorial is not the optimal setting.

[woodfrog]

I checked your git history. That padding config was introduced by this commit with the commit message "Fix tutorial error", so it seems to be an intentional change?