Large 34B model OOM in evaluation

[kyleliang919]

This line allocates the whole model onto GPU:0 again, should have used device_map="auto" when loading. https://github.com/OpenGVLab/InternVL/blob/2577068ba16fb3c17901fb3479a48b580c99c00b/internvl_chat/eval/mmmu/evaluate_mmmu.py#L286

[czczup]

Thanks for your feedback.