Runtime discrepancy when reproducing GAD-NR_inj_cora.ipynb on RTX 3080 #9

Hello,

I am trying to reproduce your GAD-NR experiments using the official GAD-NR_inj_cora.ipynb notebook. However, I noticed a large discrepancy in training time: on my setup (RTX 3080, CUDA 12.1, PyTorch 2.2.1) a single run takes about 16 hours, whereas in your notebook you mention that it only takes roughly 2 hours.

Could you please clarify:

1. Which hardware and PyTorch / PyG / torch-scatter / torch-sparse versions you used for the 2-hour runtime?
2. Whether there are any environment or configuration details (e.g., batch size, neighbor sampling, printing frequency) that are important for reproducing the reported speed?
3. Any tips to ensure the notebook runs on GPU efficiently as in your reported runtime?

Thank you very much for your help!