Cannot load torchscript model in C++

[bybeye]

Hi,

I tried to use C++ to load a torchscript geometric model with the code "pytorch_geometric/examples/cpp", and have followed the provided instructions to install pytorch_scatter and pytorch_sparse. The program has been successfully compiled, but I encountered an issue when attempting to run ./hello-world using the model generated by save_model.py.

terminate called after throwing an instance of 'torch::jit::ErrorReport'
  what():
Unknown builtin op: pyg::index_sort.
Could not find any similar ops to pyg::index_sort. This op may not exist or may not be currently supported in TorchScript.
:
  File "code/__torch__/pyg_lib/ops.py", line 11
    _1 = (_2, _3)
  else:
    _4, _5 = ops.pyg.index_sort(inputs, max_value)
             ~~~~~~~~~~~~~~~~~~ <--- HERE
    _1 = (_4, _5)
  return _1
'index_sort' is being compiled since it was called from 'index_sort'
  File "/home/edward/.local/lib/python3.9/site-packages/torch_sparse/utils.py", line 21
    if not torch_sparse.typing.WITH_INDEX_SORT:  # pragma: no cover
        return inputs.sort()
    return pyg_lib.ops.index_sort(inputs, max_value)
           ~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
Serialized   File "code/__torch__/torch_sparse/utils.py", line 3
def index_sort(inputs: Tensor,
    max_value: Optional[int]=None) -> Tuple[Tensor, Tensor]:
  _0 = __torch__.pyg_lib.ops.index_sort(inputs, max_value, )
  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
  return _0
'index_sort' is being compiled since it was called from 'SparseStorage.__init__'
Serialized   File "code/__torch__/torch_sparse/storage.py", line 25
    is_sorted: bool=False,
    trust_data: bool=False) -> NoneType:
    _0 = __torch__.torch_sparse.utils.index_sort
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
    _1 = uninitialized(int)
    _2 = uninitialized(Tensor)
'SparseStorage.__init__' is being compiled since it was called from 'is_sparse'
Serialized   File "code/__torch__/torch_geometric/utils/sparse.py", line 6
    _1 = True
  else:
    _2 = isinstance(src, __torch__.torch_sparse.tensor.SparseTensor)
                    ~~~ <--- HERE
    _1 = _2
  return _1
'is_sparse' is being compiled since it was called from 'GINConvJittable_ec52ab._check_input__0'
Serialized   File "code/__torch__/GINConvJittable_ec52ab.py", line 31
    edge_index: Tensor,
    size: Optional[Tuple[int, int]]) -> List[Optional[int]]:
    _2 = __torch__.torch_geometric.utils.sparse.is_sparse
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
    _3 = "Flow direction \"target_to_source\" is invalid for message propagation via `torch_sparse.SparseTensor` or `torch.sparse.Tensor`. If you really want to make use of a reverse message passing flow, pass in the transposed sparse tensor to the message passing module, e.g., `adj_t.t()`."
    _4 = "Expected \'edge_index\' to be of integer type (got \'{}\')"``

I have tried to use pytorch_geometric 2.3.1 and 2.4 and get the same error, but a ResNet model works well.

Environment

• PyTorch version: 2.1
• OS: WSL Ubuntu
• Python version: 3.9
• CUDA/cuDNN version: 12.2
• How you installed PyTorch and PyG (conda, pip, source): pip
• Any other relevant information (e.g., version of torch-scatter): master

[rusty1s]

Two options:

• You disable pyg_lib when converting your model to a torchscript program
• You also ensure to build against pyg-lib when loading your model in C++

[bybeye]

Thank you very much for your swift response. After uninstalling pyg_lib in Python and re-saving the model, it can be loaded in C++ now.

[qiuqiouba]

Two options:

• You disable pyg_lib when converting your model to a torchscript program
• You also ensure to build against pyg-lib when loading your model in C++

Can you explain how to build pyg-lib in C++? According to examples/cpp/README.md, I think that pyg-lib works in C++ and only needs to configure the C++APIs of TorchScatter and TorchSparse. Is this correct?

[rusty1s]

I must admit that I have never tried to run PyG with pyg-lib in C++, but you would need to build it as described here: https://github.com/pyg-team/pyg-lib/blob/master/.github/CONTRIBUTING.md