result of bert positional embed changed in 2.8

[jianyizh]

🐛 Describe the bug

when install xpu build pytorch but run on cpu, I have the following error. It will not happen on cuda

import math
import torch

div_term = (torch.arange(0, 768, 2).float() * -(math.log(10000.0) / 768)).exp()
torch.save(div_term,"d0.pt")

Compare the result of 2.7 release and 2.8rc, they are different, can cause torchbench bert accuracy failed.

>>> import torch
>>> a=torch.load("d2.7.pt")
>>> b=torch.load("d2.8.pt")
>>> a==b
tensor([ True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True, **False**,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True, **False**,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True, **False**,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True, False,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True, False,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True, False,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True, False,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True, False,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True, False,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True,  True,  True,  True,  True,  True,  True,
         True,  True,  True,  True])

Versions

it should be a compiler issue when upgrade from 2025.0 to 2025.1

[jianyizh]

The issue is the exp() op. If I build without mkl_xpu, and only source compiler and ccl, the results are same. If I source mkl, they will be different. I suppose it's a regression from mkl 2025.1 and 2025.0 @CuiYifeng

[jianyizh]

aten/src/ATen/cpu/vml.h IMPLEMENT_VML_MKL(exp, Exp)

[CuiYifeng]

Need to check oneMKL with C++ case.

[CuiYifeng]

Accuracy change of vmsExp has been reproduced with MKL C++ case.

[CuiYifeng]

oneMKL mentioned that such accuracy change of vmsExp is expected.

[jianyizh]

I still cannot reproduce the model fail under my local env.

[chuanqi129]

@jianyizh will continue to investigate the root cause in inductor and triton
@CuiYifeng check the mkl-static link issue in PyTorch CPU part