Segfault on unloading a model

[seanxhughes]

Hi,
I'm using pytorch compiled from source and with this very simple program which successfully loads a pre-trained model, then terminates I get a segfault in the sequence of destructors - anyone seen this before?

Code

#include <iostream>
#include <torch/script.h> 
#include <memory>

int main(int argc, char **argv)
{

    try {
        std::cout << "Attempting to load " << argv[1] << "\n";
        // Deserialize the ScriptModule from a file using torch::jit::load().
        auto module = torch::jit::load(argv[1]);

        std::cout << "Model loaded\n";
    }
    catch (const c10::Error& e) {
        std::cerr << "Error loading Pytorch model " << e.msg() << "\n";
        return -1;
    }

    return 0;
}

Stack trace

#0 0x000055555555a27a in c10::IValue::isTensor (this=0x80) at /usr/local/libtorch/include/ATen/core/ivalue.h:394
#1 0x000055555555ac77 in c10::IValue::destroy (this=0x80) at /usr/local/libtorch/include/ATen/core/ivalue.h:991
#2 0x000055555555a24c in c10::IValue::~IValue (this=0x80, __in_chrg=) at /usr/local/libtorch/include/ATen/core/ivalue.h:197
#3 0x0000555555561419 in std::_Destroyc10::IValue (__pointer=0x80) at /usr/include/c++/9/bits/stl_construct.h:98
#4 0x00005555555604a2 in std::_Destroy_aux::__destroyc10::IValue* (__first=0x80, __last=0x21) at /usr/include/c++/9/bits/stl_construct.h:108
#5 0x000055555555f324 in std::_Destroyc10::IValue* (__first=0x80, __last=0x21) at /usr/include/c++/9/bits/stl_construct.h:137
#6 0x000055555555d79d in std::_Destroy<c10::IValue*, c10::IValue> (__first=0x80, __last=0x21) at /usr/include/c++/9/bits/stl_construct.h:206
#7 0x000055555555c151 in std::__cxx1998::vector<c10::IValue, std::allocatorc10::IValue >::~vector (this=0x555556b2e6d0, __in_chrg=) at /usr/include/c++/9/bits/stl_vector.h:677
#8 0x000055555555a220 in std::__debug::vector<c10::IValue, std::allocatorc10::IValue >::~vector (this=0x555556b2e6b8, __in_chrg=) at /usr/include/c++/9/debug/vector:220
#9 0x000055555555dbf6 in c10::ivalue::Object::~Object (this=0x555556b2e660, __in_chrg=) at /usr/local/libtorch/include/ATen/core/ivalue_inl.h:1328
#10 0x000055555555dc32 in c10::ivalue::Object::~Object (this=0x555556b2e660, _in_chrg=) at /usr/local/libtorch/include/ATen/core/ivalue_inl.h:1328
#11 0x000055555555dd6d in c10::intrusive_ptr<c10::ivalue::Object, c10::detail::intrusive_target_default_null_typec10::ivalue::Object >::reset (this=0x7fffffffdcb0) at /usr/local/libtorch/include/c10/util/intrusive_ptr.h:289
#12 0x000055555555c37c in c10::intrusive_ptr<c10::ivalue::Object, c10::detail::intrusive_target_default_null_typec10::ivalue::Object >::~intrusive_ptr (this=0x7fffffffdcb0, __in_chrg=) at /usr/local/libtorch/include/c10/util/intrusive_ptr.h:366
#13 0x000055555555b460 in torch::jit::Object::~Object (this=0x7fffffffdcb0, __in_chrg=) at /usr/local/libtorch/include/torch/csrc/jit/api/object.h:24
#14 0x000055555555b480 in torch::jit::Module::~Module (this=0x7fffffffdcb0, __in_chrg=) at /usr/local/libtorch/include/torch/csrc/jit/api/module.h:98
#15 0x00005555555589be in main (argc=2, argv=0x7fffffffddf8) at ./exe/hello_world.cpp:87

[prabhat00155]

Are you using a torchvision model? If this is a pytorch related problem, perhaps file an issue here.

[seanxhughes]

Sorry, didn't realise I posted it in the wrong forum