pytorch - 💡(How to fix) Fix Hard crash (SIGFPE) on torch.floor_divide with INT_MIN and -1, but Eager mode works normally [1 participants]

🐛 Describe the bug

When performing integer division (torch.floor_divide) on int32 tensors where the dividend is INT_MIN (-2147483648) and the divisor is -1, Eager mode handles the overflow gracefully (matching PyTorch's ATen wrap-around semantics).

However, when compiled with torch.compile(backend="inductor"), the program hard crashes (segmentation fault / SIGFPE). This happens because Inductor generates raw codegen (C++/Triton) that directly executes the integer division INT_MIN / -1, which triggers a hardware arithmetic exception and instantly kills the entire Python process without a catchable traceback.

Here is a minimal reproducible script comparing Eager mode and Compiled mode:

import torch

def f(x, y):
    return torch.floor_divide(x, y)

cf = torch.compile(f, backend="inductor")
x = torch.tensor([-2147483648], dtype=torch.int32)
y = torch.tensor([-1], dtype=torch.int32)
print(cf(x, y), flush=True)

Versions

PyTorch Version: 2.10.0+cpu

cc @chauhang @penguinwu