chenzhao / light-dist-gnn Goto Github PK

How did you set to run the ogbn-papers100M with SANCUS? I've tried shortened hidden dimension and half precision on A100 GPUs but it still goes out of GPU memory.

Now DGL requires PyTorch >= 1.9.0,
so that we need install as conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

I try to run prepare_data.py and notice that the usage of positional argument here is incorrect. The correct way to invoke the function should be

Parted_COO_Graph(self.name, i, num_parts, preprocess_for=self.preprocess_for)

Otherwise, the self.preprocess_for will be passed to the device argument of Parted_CPP_Graph.