test_update_aliases does not correctly test function outputs #2773
Description
lightning-thunder/thunder/tests/test_update_aliases.py
Lines 122 to 135 in 9e12768
This is a test for in-place ops such as Tensor.square_, Tensor.abs_, etc.
lightning-thunder/thunder/tests/test_update_aliases.py
Lines 132 to 135 in 9e12768
op.torch_reference is precisely those ops like Tensor.square_, so it returns the same tensor as its first argument. It ends up being mutated by both jitted op and eager op, and torch.testing.assert_close(actual, expected) is trivially True.
You can verify this with this patch:
diff --git a/thunder/tests/test_update_aliases.py b/thunder/tests/test_update_aliases.py
index 7cafcbec..e5fe7b8c 100644
--- a/thunder/tests/test_update_aliases.py
+++ b/thunder/tests/test_update_aliases.py
@@ -132,6 +132,7 @@ def test_update_aliases(op, device, dtype, executor, _):
j_op = executor.make_callable(op.torch_reference)
actual = j_op(*args, **sample.kwargs)
expected = op.torch_reference(*args, **sample.kwargs)
+ assert id(args[0]) == id(actual) == id(expected), f"{id(args[0]) = }, {id(actual) = }, {id(expected) = }"
torch.testing.assert_close(actual, expected, equal_nan=True)This assertion passes on all ops except for the case op = tanhshrink (which is also strange because tanhshrink is not even in-place).