Statically linking torch_scatter torch_sparse in C++ library

[ipcamit]

I am developing a C++ library that depends on torch scatter and sparse. Currently I am following the recommended route of linking against shared libtorchscatter.so etc. and it works fine.

However I would like to bundle torchscatter as git submodule or standalone source for better install ergonomics. So I added the libtorchscatter source in folder torchscatter with following line in CMakefile

add_subdirectory("torchscatter")
target_link_libraries(MyLib PRIVATE torchscatter PRIVATE ${libtorch})

I can see that it compiles without a fuss, with libtorchscatter.a being compiled and linked to my library, however when I run the test I get the dreaded error.

Could not find any similar ops to torch_sparse::ptr2ind. This op may not exist or may not be currently supported in TorchScript.

Can anyone comment or suggest a way to achieve the same? Basically I want libtorchscatter to be statically part of libMylib.so, so that it is not a runtime dependency.

[rusty1s]

Your error links to an torch-sparse issue. What happens if you include torch-sparse the same way as you include torch-scatter?

[ipcamit]

Yes. It is only when model is loaded that it creates a fuss. That does make sense. May be I can also ask this on Torch forum see what have they to say about it?

[rusty1s]

If that's possible, that would be helpful. I actually have no idea :(

[ipcamit]

I have opened the question here: https://discuss.pytorch.org/t/statically-linking-custom-ops-in-c/173839
Lets see. Thank you for your help.

[rusty1s]

What you can also try (not sure if possible), would be to build statically against torchvision and see if that works. If yes, there is definitely a problem within our libraries :(

[ipcamit]

OK. I have found C++ examples of torchvision models should be simple enough. Will update once I have the answer