Phlippe_resnet bf16 got fail_accuracy

[mengfei25]

🐛 Describe the bug

torchbench_bfloat16_training
xpu train phlippe_resnet
E0626 09:53:20.652000 139764145854272 torch/_dynamo/utils.py:1478] RMSE (res-fp64): 0.00734, (ref-fp64): 0.00047 and shape=torch.Size([]). res.dtype: torch.bfloat16, multiplier: 3.000000, tol: 0.001000
fail_accuracy

loading model: 0it [00:00, ?it/s]
loading model: 0it [00:00, ?it/s]

Versions

torch-xpu-ops: 31c4001
pytorch: 0f81473d7b4a1bf09246410712df22541be7caf3 + PRs: 127277,129120
device: PVC 1100, 803.61, 0.5.1

[retonym]

Not very large absolute error, and this model could pass if increasing tol to 5*1e-3

[retonym]

Public PR to raise tolerance: pytorch/pytorch#134192

[retonym]

this datatype is not included in Meta dashboard, not target to PT 2.6