[Bug Skip]: [New Case] segment_reduce accumulates in the input dtype, failing test_reduced_precision_accumulation for float16/bfloat16 #6182

🐛 Describe the bug with skip template

Cases:
op_ut,third_party.torch-xpu-ops.test.xpu.test_segment_reductions_xpu.TestSegmentReductionsXPU,test_reduced_precision_accumulation_xpu_bfloat16
op_ut,third_party.torch-xpu-ops.test.xpu.test_segment_reductions_xpu.TestSegmentReductionsXPU,test_reduced_precision_accumulation_xpu_float16

Summary

test_reduced_precision_accumulation sums 4096 reduced-precision values and asserts the result matches a float accumulation of the same data; on XPU it is off by 3840.0 absolute (0.9375 relative), the magnitude you get when thousands of values are summed in float16/bfloat16 rather than in float. The XPU segment_reduce forward kernel keeps its running accumulator in scalar_t (src/ATen/native/xpu/sycl/SegmentReduceKernels.cpp:61, :88), so sum and mean never widen for reduced-precision inputs; the float16 and bfloat16 cases are the same defect in the two dtypes the test covers. Same root cause as the test_compare_cpu failures in the related draft, which that operator shows as a smaller drift against CPU.

ErrorLog

AssertionError: Tensor-likes are not close!

Traceback (most recent call last):
  File "/__w/torch-xpu-ops/torch-xpu-ops/pytorch/third_party/torch-xpu-ops/test/xpu/../../../../test/test_segment_reductions.py", line 581, in test_reduced_precision_accumulation
    self.assertEqual(actual, data.float().sum(0, keepdim=True).div_(4096 if reduction == "mean" else 1).to(dtype))
  File "/__w/torch-xpu-ops/torch-xpu-ops/.venv/lib/python3.12/site-packages/torch/testing/_internal/common_utils.py", line 5420, in assertEqual
    raise error_metas.pop()[0].to_error(  # type: ignore[index]
AssertionError: Tensor-likes are not close!

Mismatched elements: 1 / 1 (100.0%)
Greatest absolute difference: 3840.0 at index (0,) (up to 1e-05 allowed)
Greatest relative difference: 0.9375 at index (0,) (up to 0.016 allowed)

To execute this test, run the following from the base repo dir:
    PYTORCH_TEST_WITH_SLOW=1 python ../../test/test_segment_reductions.py TestSegmentReductionsXPU.test_reduced_precision_accumulation_xpu_bfloat16

This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

Reproduce

cd pytorch/third_party/torch-xpu-ops/test/xpu
pytest -sv test_segment_reductions_xpu.py::TestSegmentReductionsXPU::test_reduced_precision_accumulation_xpu_bfloat16

Pytorch Version

Versions

collect_env output was not captured in this run's artifact.