RuntimeError: "spmm" not implemented for 'Half'

[kfoynt]

I am getting this error from the baseline code that seems to be related to cuSPARSE not supporting half-precision https://docs.nvidia.com/cuda/cusparse/index.html
Which libraries have you used to run the baseline?

[rusty1s]

You need to update to the latest torch-sparse release.

[kfoynt]

Thanks. I did not reply earlier because I was trying to fix that. I am using torch 1.7.1 and cuda 11, I upgraded torch-sparse to 0.6.9.
However, I am now getting the following error:
`---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
in
21 seed_everything)
22
---> 23 from torch_sparse import SparseTensor
24 from torch_geometric.nn import SAGEConv, GATConv
25 from torch_geometric.data import NeighborSampler

~/berserker/lib/python3.7/site-packages/torch_sparse/init.py in
13 ]:
14 torch.ops.load_library(importlib.machinery.PathFinder().find_spec(
---> 15 f'{library}_{suffix}', [osp.dirname(file)]).origin)
16
17 if torch.cuda.is_available(): # pragma: no cover

AttributeError: 'NoneType' object has no attribute 'origin'`

I looked online a bit and found
rusty1s/pytorch_sparse#127
and
pyg-team/pytorch_geometric#2304

I seem to have the same issue. the cuda files are not created in the torch_sparse folder.
I tried using python 3.7 and 3.8, I updated all other torch_* to their latest versions, I checked if there are multiple versions of each. Unfortunately, nothing worked.

[kfoynt]

Alright. I just got it working. Probably just pure luck.

I think a combination of python 3.8 and
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.7.0+cu102.html
must have made it work.

Previously, I was trying python 3.8 or 3.7 and
pip install torch-sparse==0.6.9 -f https://pytorch-geometric.com/whl/torch-1.7.1+cu102.html