[BMG-Windows][PT2.8]Torch-xpu-ops UT got accuracy issue #1818
Description
🐛 Describe the bug
The following torch-xpu-ops uts got AssertionError: Scalars are not close! or AssertionError: Tensor-likes are not close!.
# AssertionError: Scalars are not close!
test_nn_xpu.py::TestNNDeviceTypeXPU::test_GroupNorm_general_xpu
test_reductions_xpu.py::TestReductionsXPU::test_noncontiguous_expanded__refs_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_noncontiguous_expanded_masked_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_noncontiguous_transposed__refs_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_noncontiguous_transposed_masked_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_ref_duplicate_values__refs_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_ref_duplicate_values_masked_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_ref_small_input__refs_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_ref_small_input_masked_prod_xpu_complex128
test_reductions_xpu.py::TestReductionsXPU::test_reference_masked_masked_prod_xpu_complex128
# AssertionError: Tensor-likes are not close!
FAILED test_reductions_xpu.py::TestReductionsXPU::test_ref_duplicate_values_prod_xpu_complex128
Reproducer:
conda create -n pt28_rc1 python=3.10 -y
conda activate pt28_rc1
pip install --pre torch==2.8.0 --index-url https://download.pytorch.org/whl/test/xpu
git clone https://github.com/pytorch/pytorch.git
cd pytorch
git checkout 03eb1e40f9ddf09cb9eef86ace74332e87f11a79
cd third_party
git clone https://github.com/intel/torch-xpu-ops.git
cd torch-xpu-ops
git checkout 3a9419c8bb6a98dd3e3cd473c36691fb4abeae40
pip install -r requirements.txt
pip install -r .ci\docker\requirements-ci.txt
pip install pytest-timeout
# MSVC activate
"C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build\vcvars64.bat"
cd C:\pytorch\third_party\torch-xpu-ops\test\xpu
set PYTORCH_TEST_WITH_SLOW=1
set PYTORCH_ENABLE_XPU_FALLBACK=1
pytest -v test_nn_xpu.py::TestNNDeviceTypeXPU::test_GroupNorm_general_xpu
Versions
| Machine | BMG - Win11 |
|---|---|
| stock pytorch | PyTorch: pip install --pre torch==2.8.0 --index-url https://download.pytorch.org/whl/test/xpu commit: 03eb1e40f9ddf09cb9eef86ace74332e87f11a79 |
| triton | pytorch-triton-xpu 3.4.0 |
| torch-xpu-ops | 3a9419c |
| Driver | internal driver: 32.0.101.6881 |
| Conda | python 3.10 |
| transformer | 243e186efbf7fb93328dd6b34927a4e8c8f24395 |
| Others | torchao 0.11.0 torchaudio 2.8.0a0+4e94321 torchvision 0.22.0a0+966da7e |