Installation with recent conda wheels

Thanks a lot @rusty1s for the great library and the timely updates!

The README states that torch_scatter can be installed with conda given that the user has a functional torch version >= 1.8.

Our installation workflow relied on this, but since today our builds are failing. It looks like it’s related to the recent release on the conda pyg channel.

This issue affects only installation on Linux.

Minimal working example:

conda create -n tmp  python=3.7 -y
conda activate tmp
pip3 install torch torchvision torchaudio #official torch installation instructions
python3 -c "import torch; print(torch.__version__)" # gives 1.12.0+cu102
conda install pytorch-scatter -c pyg #recommended in README

Afterwards, when importing torch_scatter I get:

  File "/home/ubuntu/miniconda3/envs/tmp/lib/python3.7/ctypes/__init__.py", line 364, in __init__
    self._handle = _dlopen(self._name, mode)
OSError: libc10_cuda.so: cannot open shared object file: No such file or directory

Notably, when installing torch_scatter via conda, torch gets downgraded:

The following NEW packages will be INSTALLED:
...
  pytorch            pkgs/main/linux-64::pytorch-1.10.2-cpu_py37hfa7516b_0
  pytorch-scatter    pyg/linux-64::pytorch-scatter-2.0.9-py37_torch_1.10.0_cu113
...

Here, pytorch uses CPU wheels but torch_scatter is looking for cuda wheels…

Debugging machine:

No LSB modules are available.
Distributor ID:	Ubuntu
Description:	Ubuntu 18.04.6 LTS
Release:	18.04
Codename:	bionic

Any idea of what is going on here @rusty1s ?

[therc]

Looks like I misread. Can you do me a favor and test with conda install pytorch-scatter -c pyg -c conda-forge?

On a related tangent… when you rebuild packages in the Conda channel, can you make sure to always bump the build number? Anyone using lockfiles will have a bad time, otherwise, because there seem to have been at least four versions of e.g. pytorch-scatter-2.0.9-py38_torch_1.11.0_cpu.tar.bz2 or pytorch-scatter-2.0.9-py38_torch_1.11.0_cu113.tar.bz2, with identical filenames, but all with different checksums.

[rusty1s]

One last option would be

conda install pytorch-scatter -c pyg -c pytorch -c nvidia -c conda-forge