Polygamma: UT failure

[fengyuan14]

Triage is done

• test_reference_numerics_normal_polygamma_polygamma_n_1_xpu_float16: cpu issue torch cpu float16 range is not aligned with scipy on polygamma. pytorch/pytorch#132386
• test_reference_numerics_normal_polygamma_polygamma_n_2_xpu_float16: cpu issue torch cpu float16 range is not aligned with scipy on polygamma. pytorch/pytorch#132386
• test_reference_numerics_normal_polygamma_polygamma_n_3_xpu_float16: cpu issue torch cpu float16 range is not aligned with scipy on polygamma. pytorch/pytorch#132386
• test_reference_numerics_normal_polygamma_polygamma_n_4_xpu_float16: cpu issue torch cpu float16 range is not aligned with scipy on polygamma. pytorch/pytorch#132386

🐛 Describe the bug

Please evaluate inf in the error log, whether it is caused by numerical error when calculating a extremal large number. 4 failures,

=========================== short test summary info ============================
FAILED test_unary_ufuncs_xpu.py::TestUnaryUfuncsXPU::test_reference_numerics_normal_polygamma_polygamma_n_1_xpu_float16 - AssertionError: Tensor-likes are not close!

Mismatched elements: 8 / 943593 (0.0%)
Greatest absolute difference: inf at index (9, 860) (up to 0.001 allowed)
Greatest relative difference: inf at index (9, 860) (up to 0.0012 allowed)

To execute this test, run the following from the base repo dir:
    PYTORCH_TEST_WITH_SLOW=1 python test/test_unary_ufuncs.py -k TestUnaryUfuncsXPU.test_reference_numerics_normal_polygamma_polygamma_n_1_xpu_float16

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
FAILED test_unary_ufuncs_xpu.py::TestUnaryUfuncsXPU::test_reference_numerics_normal_polygamma_polygamma_n_2_xpu_float16 - AssertionError: Tensor-likes are not close!

Mismatched elements: 8 / 943593 (0.0%)
Greatest absolute difference: inf at index (9, 860) (up to 0.001 allowed)
Greatest relative difference: inf at index (9, 860) (up to 0.0012 allowed)

To execute this test, run the following from the base repo dir:
    PYTORCH_TEST_WITH_SLOW=1 python test/test_unary_ufuncs.py -k TestUnaryUfuncsXPU.test_reference_numerics_normal_polygamma_polygamma_n_2_xpu_float16

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
FAILED test_unary_ufuncs_xpu.py::TestUnaryUfuncsXPU::test_reference_numerics_normal_polygamma_polygamma_n_3_xpu_float16 - AssertionError: Tensor-likes are not close!

Mismatched elements: 8 / 943593 (0.0%)
Greatest absolute difference: inf at index (9, 860) (up to 0.001 allowed)
Greatest relative difference: inf at index (9, 860) (up to 0.0012 allowed)

To execute this test, run the following from the base repo dir:
    PYTORCH_TEST_WITH_SLOW=1 python test/test_unary_ufuncs.py -k TestUnaryUfuncsXPU.test_reference_numerics_normal_polygamma_polygamma_n_3_xpu_float16

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
FAILED test_unary_ufuncs_xpu.py::TestUnaryUfuncsXPU::test_reference_numerics_normal_polygamma_polygamma_n_4_xpu_float16 - AssertionError: Tensor-likes are not close!

Mismatched elements: 8 / 943593 (0.0%)
Greatest absolute difference: inf at index (9, 860) (up to 0.001 allowed)
Greatest relative difference: inf at index (9, 860) (up to 0.0012 allowed)

To execute this test, run the following from the base repo dir:
    PYTORCH_TEST_WITH_SLOW=1 python test/test_unary_ufuncs.py -k TestUnaryUfuncsXPU.test_reference_numerics_normal_polygamma_polygamma_n_4_xpu_float16

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
= 4 failed, 15584 passed, 552 skipped, 2299 deselected, 16 xfailed in 131.29s (0:02:11) =

Versions

#555

[daisyden]

XPU got the similar issue as cpu backend, both cpu and xpu float16 range has gap with scipy. CPU issue is submitted pytorch/pytorch#132386 with example code.

[daisyden]

@yuchengliu1 please adjust tolerance for this case.

[daisyden]

tracked in #754