gets nan with complex dtype

[Stonepia]

🐛 Describe the bug

We get nan when the dtype is complex. The tests are:

PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal__refs_sin_xpu_complex128
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal__refs_sin_xpu_complex64
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal_nn_functional_tanhshrink_xpu_complex128
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal_nn_functional_tanhshrink_xpu_complex64
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal_sin_xpu_complex128
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal_sin_xpu_complex64
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal_sinh_xpu_complex128
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_extremal_sinh_xpu_complex64
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_large__refs_exp_xpu_complex128
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_large__refs_exp_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_large__refs_sin_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_large_exp_xpu_complex128
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_large_exp_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_large_sin_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_normal_sigmoid_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_small_acos_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_small_asin_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_small_asinh_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_small_atan_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_small_atanh_xpu_complex32
PYTORCH_TEST_WITH_SLOW=1 python test\test_unary_ufuncs.py TestUnaryUfuncsXPU.test_reference_numerics_small_sigmoid_xpu_complex32

Versions

Machine: LNL windows 11
PyTorch version: https://github.com/pytorch/pytorch/tree/90b7dcf2c5ee13b892701822f2abbc0e64f5584d
Install command: pip install --pre torch==2.6.0.dev20241202+xpu torchvision==0.20.0.dev20241202+xpu torchaudio==2.5.0.dev20241202+xpu --index-url https://download.pytorch.org/whl/nightly/xpu
torch-xpu-ops commit: bf4bab1

[Stonepia]

There should be related to bugs with the compiler. We will discuss on how to co-work with the compiler team for this issue.