TypeError when passing list of tensors to C++

C++
1

I am trying to pass a TensorList to a C++ extension, as done for example in torch.nn.utils.rnn.pad_sequence. I am new to extending pytorch and have been following this tutorial.

A very basic function to test if I can pass a list of tensors could be

#include <torch/extension.h>
#include <pybind11/pybind11.h>

bool test_tensorlist(torch::TensorList x) {
  return true
}

PYBIND11_MODULE(tensorlist, m) {
    m.doc() = "Test how to pass tensor list to C++ extension.";
    m.def("test_tensorlist", &test_tensorlist, "Pass list of tensors (as torch::TensorList).")
}

Unfortunately, attempting to call this in python results in

TypeError: incompatible function arguments. The following argument types are supported: 
   1. (arg0: c10::ArrayRef<at::Tensor>) -> bool  
Invoked with: [tensor([1]), tensor([2])]

I get the exact same result if I use at::TensorList or c10::ArrayRef<at::Tensor>. Passing a single tensor via torch::Tensor or a list of integers via c10::ArrayRef<int64_t> works like charm, though.

Any ideas what I might be doing wrong?

Edits: formatting

1 Like
2

I’m having the same issue. Maybe the pybind binding for TensorList is not implemented yet. I opened an issue on GitHub here: TensorList Python Binding Encountering C++ TypeError · Issue #80979 · pytorch/pytorch (github.com)

As a workaround, I was able to use the TORCH_LIBRARY bindings successfully ( pytorch/library.h at 38350acf8fddc350c893e23a8e77bed2e59f5b06 · pytorch/pytorch (github.com)). Here is a tutorial that helped: Extending TorchScript with Custom C++ Operators — PyTorch Tutorials 1.12.0+cu102 documentation

The basic function you listed above works with this approach. The downside is that there is now the extra step of calling torch.ops.load_library at runtime