Unknown builtin op: torch_scatter::scatter_max

[ryan2x]

When I tried to load TorchScript module from C++, I got the error

terminate called after throwing an instance of 'torch::jit::ErrorReport' what(): Unknown builtin op: torch_scatter::scatter_max. Could not find any similar ops to torch_scatter::scatter_max. This op may not exist or may not be currently supported in TorchScript. : File "/opt/conda/lib/python3.7/site-packages/torch_scatter/scatter.py", line 72

The model is trained using PyTorch 1.5.0 and latest pytorch geometric, and torch scatter 2.0.5, and I converted the trained model to TorchScript. I noted the document from https://pytorch-geometric.readthedocs.io/en/latest/notes/jit.html. But it did not mention how to load the custom operators such as torch_scatter::scatter_max. The serving environment is C++, and that's a separate environment from the training environment.

How can I resolve this error for TorchScript serving using C++?

[rusty1s]

That's correct. This information is currently missing in the tutorial. You need to also build the torch-scatter C++ API, see here. You can find a complete example of serving a model here.

[ryan2x]

Thanks for your prompt response. After installing torch-scatter C++ API, that is libtorchscatter.so, I can get TorchScript serving using C++ working.

[bs1119]

Thanks for your prompt response. After installing torch-scatter C++ API, that is libtorchscatter.so, I can get TorchScript serving using C++ working.

can you explain how to "installing torch-scatter C++ API" exactly? I want to do a similar thing in windows using VS2017.

[rusty1s]

See pyg-team/pytorch_geometric#1718 (comment).

[github-actions]

This issue had no activity for 6 months. It will be closed in 2 weeks unless there is some new activity. Is this issue already resolved?

[zdwmeta]

Hi @rusty1s, I am able to compile and run the following code:

#include <iostream>
#include <torch/torch.h>
#include <torchsparse/sparse.h>

int main() {
  torch::Tensor src = torch::tensor({0.5, 0.4, 0.1, 0.6});
  torch::Tensor index = torch::tensor({0, 0, 1, 1});
  std::cout << src << std::endl;
  std::cout << index << std::endl;
  std::cout << std::get<0>(
                   scatter_min(src, index, 0, torch::nullopt, torch::nullopt))
            << std::endl;
}

However, when running inference with PyG script module, I am still getting the error: Unknown builtin op: torch_scatter::scatter_min. Do you have any idea why this could still happen?

[rusty1s]

Did you take a look at https://github.com/pyg-team/pytorch_geometric/tree/master/examples/cpp? Does it work for you?