Possible to compile the neighbor function separate from full package?

[InfluenceFunctional]

I'm given to understand the neighbor list / radial graph function neighbors is very efficient, and I was hoping to test it in another pipeline I'm developing, without installing the full torchmd-net suite. Is this possible / easy to do? I'm not experienced at compiling standalone cuda+cpp modules so apologies if it is obvious.

[RaulPPelaez]

Indeed!
99% of the building time of torchmd-net comes from the neighbor list extensions. So I would not worry about wasting time by building the rest.

If you are not planning to modify the extension itself I would just conda install torchmd-net.

You can then just import this https://torchmd-net.readthedocs.io/en/latest/generated/torchmdnet.extensions.html#torchmdnet.extensions.get_neighbor_pairs_kernel
Or alternatively this wrapper: https://torchmd-net.readthedocs.io/en/latest/generated/torchmdnet.models.utils.html#torchmdnet.models.utils.OptimizedDistance

Building torch extensions is really tricky, so I would start by using the conda builds. If you want to build locally to fiddle with the code you can also use conda to get the depencencies: https://torchmd-net.readthedocs.io/en/latest/installation.html#install-from-source
Even so, if you want to take it for your own project, I would hijack torchmd-net's already in place extension building system (consisting on setup.py + extensions/init.py + extensions.cpp) . It took a lot of sweat to make it work hehe.

[InfluenceFunctional]

Thanks very much for getting back to me on this! Good to know I'm not alone in finding it complex/difficult.
Will probably just install the full torchmd-net as you suggest.