[Inductor] [CPU] Accuracy failure in huggingface model XLNetLMHeadModel on WW03.2

[yudongsi]

🐛 Describe the bug

Compare with the ww02.4, there is a accuracy failure on Huggingface model XLNetLMHeadModel on TorchInductor CPU Performance Dashboard on ww03.2:

WW03.2 SW info:

SW	Nightly commit	Master/Main commit
Pytorch	868797a	d41b5d7
Torchbench	/	bbdc777
torchaudio	ecc2781	4a037b0
torchtext	112d757	c7cc5fc
torchvision	ac06efe	35f68a0
torchdata	049fb62	c0934b9
dynamo_benchmarks	fac4361	/

WW02.4 SW info:

SW	Nightly commit	Master/Main commit
Pytorch	466be3f	364f526
Torchbench	/	bbdc777
torchaudio	ecc2781	4a037b0
torchtext	112d757	c7cc5fc
torchvision	ac06efe	35f68a0
torchdata	049fb62	c0934b9
dynamo_benchmarks	fac4361	/

Error logs

cpu  eval  XLNetLMHeadModel                   [2023-01-16 07:53:30,105] torch._dynamo.utils: [ERROR] RMSE (res-fp64): 0.01949, (ref-fp64): 0.00000 and shape=torch.Size([])
[2023-01-16 07:53:30,106] torch._dynamo.utils: [ERROR] Accuracy failed for key name loss
FAIL

Minified repro

python -m torch.backends.xeon.run_cpu --core_list 0 --ncores_per_instance 1 benchmarks/dynamo/huggingface.py --accuracy --float32 -dcpu --inductor  --no-skip --dashboard --only XLNetLMHeadModel --batch_size 1 --threads 1

[Guobing-Chen]

Fixed with PR#92299