bug.cc

/* Build & run with
#   conda create -n bugenv python=3.7
#   conda activate bugenv
#   pip install 'torch==1.3.0'
#   CXX=c++ python3 setup.py build develop
#   python run.py
*/

#include <torch/extension.h>

class Cat {
 public:
  at::Tensor cat(std::vector<at::Tensor> tensors, int dim) {
    return at::cat(tensors, dim);
  }

  at::Tensor cat_with_catch_c10(std::vector<at::Tensor> tensors, int dim) {
    try {
      return cat(tensors, dim);
    } catch (const c10::IndexError& e) {
      std::cout << "c10::IndexError with " << e.what() << std::endl;
    } catch (...) {
      int status;
      std::cout << "... exception! Let's see which kind: "
                << abi::__cxa_demangle(
                       abi::__cxa_current_exception_type()->name(), 0, 0,
                       &status)
                << std::endl;
    }
    return torch::zeros(0);
  }

  at::Tensor cat_with_catch_exception(std::vector<at::Tensor> tensors,
                                      int dim) {
    try {
      return cat(tensors, dim);
    } catch (const std::exception& e) {
      std::cout << "std::exception with " << e.what() << std::endl;
    } catch (...) {
      int status;
      std::cout << "... exception! Let's see which kind: "
                << abi::__cxa_demangle(
                       abi::__cxa_current_exception_type()->name(), 0, 0,
                       &status)
                << std::endl;
    }
    return torch::zeros(0);
  }

 private:
  std::vector<torch::Tensor> results_;
};

PYBIND11_MODULE(tensorbug, m) {
  py::class_<Cat>(m, "Cat")
      .def(py::init<>())
      .def("cat", &Cat::cat)
      .def("cat_with_catch_c10", &Cat::cat_with_catch_c10)
      .def("cat_with_catch_exception", &Cat::cat_with_catch_exception);
}

run.py

# Build & run with
#   conda create -n bugenv python=3.7
#   conda activate bugenv
#   pip install 'torch==1.3.0'
#   CXX=c++ python3 setup.py build develop
#   python run.py 0   # or 1 or 2

# Behavior tested on
#   OS: Mac OSX 10.14.3
#   GCC version: Could not collect
#   CMake version: version 3.12.2
#   Python version: 3.7
# with:
#   pip install 'torch==1.3.0'
#   pip install 'torch==1.2.0'
#   pip install 'torch==1.1.0'
#   pip install 'torch==1.0.0'  (needs s/c10::IndexError/c10::Error/)
#
# WORKS on
#   OS: Ubuntu 18.04.2 LTS
#   GCC version: (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0
#   CMake version: version 3.12.2

import sys

import torch
import tensorbug

catter = tensorbug.Cat()

codepath = int(sys.argv[1]) if len(sys.argv) > 1 else 2

if codepath == 0:
    ## Just plain breaks:
    catter.cat([torch.zeros(1)], 10)

if codepath == 1:
    ## Semi-good: Catches exception but breaks anyway.
    ## Can at least be debugged.
    catter.cat_with_catch_c10([torch.zeros(1)], 10)

if codepath == 2:
    ## Worst: Does not catch std::exception.
    ## Non-portable inspection shows it's an index error.
    catter.cat_with_catch_exception([torch.zeros(1)], 10)

setup.py

# Build with
#   CXX=c++ python3 setup.py build develop

import setuptools
import sys

from torch.utils import cpp_extension

extra_compile_args = []
extra_link_args = []

if sys.platform == "darwin":
    extra_compile_args += ["-stdlib=libc++", "-mmacosx-version-min=10.14"]
    extra_link_args += ["-stdlib=libc++"]


tensorbug = cpp_extension.CppExtension(
    name="tensorbug",
    sources=["bug.cc"],
    language="c++",
    extra_compile_args=["-std=c++17"] + extra_compile_args,
    extra_link_args=extra_link_args,
)


setuptools.setup(
    name="tensorbug",
    ext_modules=[tensorbug],
    cmdclass={"build_ext": cpp_extension.BuildExtension},
)