Inconsistent Hash of IValue between aten/src/ATen/core/ivalue.cpp and aten/src/ATen/core/Dict_inl.h

[HunterTracer]

🐛 Describe the bug

Recently, I'm using libtorch 1.13.0 using C++ in VS 2022. I want to create a torch::Dict, which maps tuple<string, string, string> to tuple<string, string, string>. After I create the tuple key, I use torch::IValue::hash(key) and this works. However, when I want to insert (key, value) into dict, the program raises error in DictKeyHash::operator()(const IValue& ivalue) (the function is in aten/src/ATen/core/Dict_inl.h). It says can't hash IValues.

The reason is that DictKeyHash::operator() only supports to hash int, string, double, complex double, bool, tensor and device. However, the funciton torch::IValue::hash in aten/src/ATen/core/ivalue.cpp supports None, Bool, Double, Tensor, Storage, Int, SymInt, SymFloat, String, Tuple, and Device. There is a conflict between two hash function. The first one cannot hash None, Storage, SymInt, SymFloat, Tuple, and the second one cannot hash complex double.

Is it a bug? Can these hash functions be consistent? Thanks!

Here's my code.

`#include
#include <torch/torch.h>

int main() {
torch::TupleTypePtr pt = torch::TupleType::create({torch::StringType::get(), torch::StringType::get(), torch::StringType::get()});
std::cout << pt->str() << "\n";

c10::Dict<torch::IValue, torch::IValue> dict(pt, pt);
torch::IValue key(std::make_tuple("one", "one", "one"));
std::cout << torch::IValue::hash(key) << "\n";
dict.insert(key, key); // <- will fail in DictKeyHash::operator()
return 0;

}
`

Versions

Windows 10, Visual Studio 2022, C++, libtorch 1.13.0 with cuda

cc @EikanWang @jgong5 @wenzhe-nrv @sanchitintel @jbschlosser

[malfet]

Why do you want/need to use Dict rather than std::unordered_map?