use GATv2Conv or GATConv, but RuntimeError: Not compiled with CUDA support occured

[yayuanzi8]

The error message is :

File "/home/cxj/anaconda3/envs/code_graph_nn_pyg/lib/python3.7/site-packages/torch_geometric/nn/conv/gatv2_conv.py", line 156, in forward
out = self.propagate(edge_index, x=(x_l, x_r), size=size)
File "/home/cxj/anaconda3/envs/code_graph_nn_pyg/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 309, in propagate
out = self.message(**msg_kwargs)
File "/home/cxj/anaconda3/envs/code_graph_nn_pyg/lib/python3.7/site-packages/torch_geometric/nn/conv/gatv2_conv.py", line 183, in message
alpha = softmax(alpha, index, ptr, size_i)
RuntimeError: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript (most recent call last):
File "/home/cxj/anaconda3/envs/code_graph_nn_pyg/lib/python3.7/site-packages/torch_scatter/scatter.py", line 160, in softmax
return scatter_min(src, index, dim, out, dim_size)[0]
elif reduce == 'max':
return scatter_max(src, index, dim, out, dim_size)[0]
~~~~~~~~~~~ <--- HERE
else:
raise ValueError
File "/home/cxj/anaconda3/envs/code_graph_nn_pyg/lib/python3.7/site-packages/torch_scatter/scatter.py", line 72, in scatter_max
out: Optional[torch.Tensor] = None,
dim_size: Optional[int] = None) -> Tuple[torch.Tensor, torch.Tensor]:
return torch.ops.torch_scatter.scatter_max(src, index, dim, out, dim_size)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
RuntimeError: Not compiled with CUDA support

Version info：

torch 1.8.1
torch-cluster 1.5.9
torch-geometric 2.0.2
torch-scatter 2.0.9
torch-sparse 0.6.12
torch-spline-conv 1.2.1

cuda version is 10.1

[rusty1s]

Can you say something about how you did install torch-scatter? It looks like it only did install the CPU version in your case.

[yayuanzi8]

I can use scatter_mean, scatter_add using GPU, but only can't run scatter_max

[rusty1s]

Yes, this is to be expected as we use torch.scatter_add_ internally for these variants. Can you try to uninstall torch-scatter and post the installation log of:

pip uninstall torch-scatter
pip install --no-cache-dir --verbose --no-index torch-scatter -f https://data.pyg.org/whl/torch-1.8.1+cu101.html

[yayuanzi8]

After reinstall torch-scatter by running pip install --no-cache-dir --verbose --no-index torch-scatter -f https://data.pyg.org/whl/torch-1.8.1+cu101.html，the problem solved，now I can use GATConv and GATv2Conv. The log is:

Using pip 21.2.2 from /home/cxj/anaconda3/envs/code_graph_nn_pyg/lib/python3.7/site-packages/pip (python 3.7)
Looking in links: https://data.pyg.org/whl/torch-1.8.1+cu101.html
Collecting torch-scatter
Downloading https://data.pyg.org/whl/torch-1.8.0%2Bcu101/torch_scatter-2.0.8-cp37-cp37m-linux_x86_64.whl (7.9 MB)
|████████████████████████████████| 7.9 MB 870 kB/s
Installing collected packages: torch-scatter
Successfully installed torch-scatter-2.0.8

I found the version is 2.0.8, instead of 2.0.9. But why？

[rusty1s]

Ah, you were using an older PyTorch version. For PyTorch 1.8.1, binaries of torch-scatter do only exist till version 2.0.8, see here. This shouldn't be a problem at all.

[yayuanzi8]

Thank you for your help.