[MLT Windows] multihead_attention float16 got accuracy issue

[daisyden]

🐛 Describe the bug

FAILED test_native_mha_xpu.py::TestMHADeviceTypeXPU::test_native_multihead_attention_xpu_float16
================================== FAILURES ===================================
______ TestMHADeviceTypeXPU.test_native_multihead_attention_xpu_float16 _______
Traceback (most recent call last):
File "C:\Users\gta\penghuic\pytorch\third_party\torch-xpu-ops\test\xpu../../../../test\test_native_mha.py", line 323, in test_native_multihead_attention
self._test_multihead_attention_impl(
File "C:\Users\gta\penghuic\pytorch\third_party\torch-xpu-ops\test\xpu../../../../test\test_native_mha.py", line 256, in _test_multihead_attention_impl
torch.testing.assert_close(ypt, ynpt.to(torch.float32), atol=1e-3, rtol=1e-3)
File "C:\Users\gta\Miniforge3\envs\ut-py310\lib\site-packages\torch\testing_comparison.py", line 1530, in assert_close
raise error_metas[0].to_error(msg)
AssertionError: Tensor-likes are not close!

Mismatched elements: 416 / 8192 (5.1%)
Greatest absolute difference: 0.0026940107345581055 at index (2, 7, 62) (up to 0.001 allowed)
Greatest relative difference: 10.974992752075195 at index (10, 4, 51) (up to 0.001 allowed)

To execute this test, run the following from the base repo dir:
python test\test_native_mha.py TestMHADeviceTypeXPU.test_native_multihead_attention_xpu_float16

FAILED test_native_mha_xpu.py::TestMHADeviceTypeXPU::test_native_multihead_encoder_decoder_attention_xpu_float16

================================== FAILURES ===================================
_ TestMHADeviceTypeXPU.test_native_multihead_encoder_decoder_attention_xpu_float16 _
Traceback (most recent call last):
File "C:\Users\gta\penghuic\pytorch\third_party\torch-xpu-ops\test\xpu../../../../test\test_native_mha.py", line 309, in test_native_multihead_encoder_decoder_attention
self._test_multihead_attention_impl(
File "C:\Users\gta\penghuic\pytorch\third_party\torch-xpu-ops\test\xpu../../../../test\test_native_mha.py", line 256, in _test_multihead_attention_impl
torch.testing.assert_close(ypt, ynpt.to(torch.float32), atol=1e-3, rtol=1e-3)
File "C:\Users\gta\Miniforge3\envs\ut-py310\lib\site-packages\torch\testing_comparison.py", line 1530, in assert_close
raise error_metas[0].to_error(msg)
AssertionError: Tensor-likes are not close!

Mismatched elements: 454 / 8192 (5.5%)
Greatest absolute difference: 0.0034734010696411133 at index (1, 3, 48) (up to 0.001 allowed)
Greatest relative difference: 1.2274675369262695 at index (1, 6, 26) (up to 0.001 allowed)

To execute this test, run the following from the base repo dir:
python test\test_native_mha.py TestMHADeviceTypeXPU.test_native_multihead_encoder_decoder_attention_xpu_float16

Versions

nightly 20241230