why I use the default configuration file and only change the batch size to 8 ,I can't train with four 3080 Gpus ?

Hello,why I use the default configuration file and only change the batch size to 8 ,I can't train with four 3080 Gpus ?

1. Full runnable code or full changes you made:
```
/home/flycar/nj/detectron2/projects/Panoptic-DeepLab/configs/Cityscapes-PanopticSegmentation/Base-PanopticDeepLab-OS16.yaml
```

2. What exact command you run:python train_net.py --config-file configs/Cityscapes-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024_dsconv.yaml --num-gpus 4
3. __Full logs__ or other relevant observations:

## Environment:
The environments I configured were: ubuntu18.04, cuda 11.1, pytorch 1.10
Paste the output of the following command:
```
python train_net.py --config-file configs/Cityscapes-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024_dsconv.yaml --num-gpus 4

```

Start by reading the referenced Base-PanopticDeepLab-OS16.yaml and the command's panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024_dsconv.yaml configuration, then reproduce the command in the stated Ubuntu, CUDA, and PyTorch environment. Inspect the complete training logs and determine why four GPUs cannot train with batch size 8; done means the cause is identified and a reproducible configuration or documented limitation is established.

Written by the indexing model from the issue text.