Wrong accuracy of aten::amax after update driver from LTS to LTS2

[etaf]

🐛 Describe the bug

The Inductor test case "python test/inductor/test_torchinductor_opinfo.py test/inductor/test_torchinductor_opinfo.py TestInductorOpInfoXPU.test_comprehensive_masked_amax_xpu_int64" failed when trying to update driver from LTS to LTS2, and if we fallback the op aten::amax to cpu, the case can pass again. So the accuracy of aten::amax has accuracy regression between LTS and LTS2.

Reproduce:
LTS:
python test/inductor/test_torchinductor_opinfo.py TestInductorOpInfoXPU.test_comprehensive_masked_amax_xpu_int64 : pass

LTS2:
python test/inductor/test_torchinductor_opinfo.py TestInductorOpInfoXPU.test_comprehensive_masked_amax_xpu_int64 : fail

PYTORCH_DEBUG_XPU_FALLBACK=1 PYTORCH_XPU_FALLBACK_OP=amin python test/inductor/test_torchinductor_opinfo.py TestInductorOpInfoXPU.test_comprehensive_masked_amax_xpu_int64: pass

Versions

Pytorch: main or nightly.

[CuiYifeng]

Reported to IGC. Guilty driver version has been found.

[CuiYifeng]

The fixing of driver is expected to be available in PT2.10.

[mengfei25]

Passed with new driver.