OOM after Postprocessing on video datasets

[xumingze0308]

I evaluated Oryx-34B on VideoMME using 8 H100 GPUs. The "Model Responding" works well but got OOM (NCCL WARN Cuda failure 2 'out of memory') after the Postprocessing was completed.

I tried #4 but it seems to only use 1 GPU. Is there any suggestion on this problem? I feel maybe metric calculation or gathering causes OOM and am curious if this can be optimized.

Thank you!

[xumingze0308]

This issue can be solved by this PR.