[UT] torch-xpu-ops/test/profiling/reproducer.missing.gpu.kernel.time.py core dumped

[RUIJIEZHONG66166]

🐛 Describe the bug

Description

Test case: torch-xpu-ops/test/profiling/reproducer.missing.gpu.kernel.time.py now coredump with latest nightly wheel.
The error reported on the PVC is as follows:

 import pynvml  # type: ignore[import]
======================================
Segmentation fault from GPU at 0xff00000003609000, ctx_id: 1 (CCS) type: 0 (NotPresent), level: 1 (PDE), access: 1 (Write), banned: 1, aborting.
Segmentation fault from GPU at 0xff00000003609000, ctx_id: 1 (CCS) type: 0 (NotPresent), level: 1 (PDE), access: 1 (Write), banned: 1, aborting.
Abort was called at 288 line in file:
./shared/source/os_interface/linux/drm_neo.cpp
Aborted (core dumped)

The error reported on the B70 is as follows:

======================================
Traceback (most recent call last):
  File "/root/ruijie/torch-xpu-ops/test/profiling/reproducer.missing.gpu.kernel.time.py", line 73, in <module>
    maxUnpool2d(shape, dtype, channels_last, backward=backward)
  File "/root/ruijie/torch-xpu-ops/test/profiling/reproducer.missing.gpu.kernel.time.py", line 41, in maxUnpool2d
    x_dpcpp = output.to(device="xpu", dtype=dtype)
RuntimeError: level_zero backend failed with error: 20 (UR_RESULT_ERROR_DEVICE_LOST)

I have did some triages, then found:
torch-2.14.0.dev20260811+xpu: This case runs correctly.
torch-2.15.0.dev20260812+xpu: You will encounter this problem.

Reproduction

conda create -yn test cmake ninja pillow python=3.12
conda activate test
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu
python torch-xpu-ops/test/profiling/reproducer.missing.gpu.kernel.time.py

Versions

torch version: pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu
Device: B70/PVC linux