multi-gpu support

[GritLs]

Thank you for your outstanding work on this project and for making it open source!

I would like to inquire if multi-GPU training is supported. I noticed a switch on line 78 of run.py that controls whether to use multi-GPU. However, when I set this option to true and rerun the script, the process seems to get stuck, and the terminal outputs the following message:

Could you please assist me with this issue? Thank you!

[hvlgo]

Please provide the modified scripts, or you can refer to this. I guess it's most likely because use_multi_gpu was used, but torchrun wasn't used to start.