It is inside the PyTorch source code. This patch was already applied to the PyTorch wheels (for PyTorch >= 1.1.0) if you are using one of the pre-built PyTorch wheels from this topic:
https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-8-0-now-available/72048
If you are still getting the error, try reducing the --batch-size when running train_ssd.py