pytorch - 💡(How to fix) Fix torch.compile (inductor/aot_eager) produces nondeterministic / incorrect results on model with `add_`, `clamp_`, and `squeeze` [2 comments, 3 participants]

🐛 Describe the bug

When using torch.compile with inductor and aot_eager backends on a function including add_, clamp_, and squeeze, results are nondeterministic and depend on the execution order.

Here is the reproducible code and the actual problem:

import os
os.environ["TORCHINDUCTOR_FORCE_DISABLE_CACHES"] = "1"
import torch

def fn(x, y):
    x.add_(1)
    y.clamp_(min=0)
    return x.squeeze(1, 2).clone()

backend="inductor"

def run1():
    cfunc = torch.compile(fn, backend=backend)
    a = torch.full((1, 1, 1), -10.0)
    b = a.view(-1).clone()
    out1 = cfunc(a, b)
    print("out1: ", out1)  # expected: -9, actual: -9

def run2():
    cfunc = torch.compile(fn, backend=backend)
    a = torch.full((1, 1, 1), -10.0)
    b = a.view(-1)
    out2 = cfunc(a, b)
    print("out2: ", out2)  # expected: 0, actual: inductor gives 1, aot_eager gives -9

# if we directly execute run2(), out2's actual result is 0
# but if we execute run1() before run2(), out2's actual result becomes incorrect
run1()
run2()

Expected behavior: run1() -> -9 run2() -> 0

Actual behavior:

• if execute run2() alone, both inductor and aot_eager output 0 (correct).
• if first execute run1(), then execute run2(), inductor gives 1, aot_eager gives -9.

Error logs

No response

Versions

[pip3] numpy==2.4.2 [pip3] torch==2.12.0.dev20260328+cpu [pip3] torchvision==0.26.0.dev20260223+cpu [pip3] triton==3.6.0 [conda] numpy 2.4.2 pypi_0 pypi [conda] torch 2.12.0.dev20260328+cpu pypi_0 pypi [conda] torchvision 0.26.0.dev20260223+cpu pypi_0 pypi [conda] triton 3.6.0 pypi_0 pypi

cc @chauhang @penguinwu